Anthropic's consumer and team AI assistant apps for web, iOS, Android, and desktop, covering chat, projects, artifacts, research, memory, and connectors.
Software capabilities Identity & access Account deletion Lets a user delete their Claude account and holds the email address for 30 days before it can be reused for a new account. Account login methods Signs a user in with an emailed login code, a Google account, or their organization single sign-on. Active session management Lists every browser and device where the account is signed in with device, location, and last use, and lets the user end any session or sign out everywhere. Age assurance Requires a user flagged as possibly under 18 to confirm they are 18 or over through Yoti facial age estimation, ID upload, or a Yoti digital ID before the account is reinstated. Audit logs (Claude) Lets admins export a log of security- and access-relevant events across the organization. Claude for Government Runs Claude inside a FedRAMP High authorized environment isolated from commercial tenants, routing inference through Google Cloud Vertex AI and gating Opus-class models by seat tier. Custom data retention controls Lets admins set how long the organization retains chats, projects, and private or shared artifacts, as a fixed window per data type or indefinitely. HIPAA-ready Enterprise plans Lets an Enterprise primary owner accept Anthropic's Business Associate Agreement and enable a HIPAA configuration for the organization, after which chat, projects, artifacts, research, skills, and file creation are eligible services for processing protected health information. Identity verification Prompts a user to confirm who they are with a government-issued photo ID and a live selfie through Persona before certain capabilities become available. Incognito chats Lets users have chats that are excluded from chat history and account personalization. IP allowlisting Validates the source IP of every authenticated request against an Enterprise organization CIDR allowlist and blocks the rest. Just-in-time provisioning Creates a member account and assigns its role from identity-provider attributes on first SSO login, without a directory sync. Organization discovery Lets someone signing up with a work email find their company existing organization and request to join, or be added automatically. Parent organizations Holds domain verification, SSO configuration, and provisioning settings at a parent entity shared by several Claude and Console organizations, which join through a merge proposal. Personal account migration and domain claiming Moves a personal Free, Pro, or Max account chats, projects, and artifacts into a Team or Enterprise organization, either started by the user or claimed by an admin on a verified domain. Phone number verification Requires a phone number from a supported location and an SMS code before a new account can be used. SCIM provisioning (Claude) Automatically provisions and deprovisions Claude user accounts from an external identity provider via SCIM. Session duration controls Lets Enterprise and Console admins cap how long a session stays valid before every member must sign in again. SSO and domain capture Lets an organization authenticate members through an enterprise identity provider and capture accounts under its verified email domains. Tenant restrictions Blocks access to Claude from a corporate network for any organization not on the allowed list, by validating a header the network proxy injects. Trusted Devices Requires a member to verify an enrolled device before viewing or steering a local Claude Code session remotely, and lets admins revoke a member's trusted devices. US-only inference Keeps an Enterprise organization model inference on servers located in the United States across every Claude app.
Admin governance Chat rating controls Lets Team and Enterprise owners turn off members ability to send thumbs-up and thumbs-down chat ratings to Anthropic. Classification banner Displays an admin-configured classification marking, with its own text, colors, and optional link, at the top of every page for every member of the organization. Custom roles Lets Enterprise admins define roles from permission sets that grant or restrict individual features, connectors, and admin areas. Default model settings Sets which Claude model new conversations start on for the whole organization or for a specific custom role. Designated support contacts Designates which groups may open tickets with Anthropic human support team while everyone else is served by AI support. Desktop extension allowlist Lets Team and Enterprise owners choose which desktop extensions members are allowed to install. Enterprise-managed connector auth Authorizes an MCP connector once for the whole organization through the identity provider so members inherit access on first login. Groups Organizes members into groups that drive role assignment, project sharing, and spend limits. Member management Lets admins invite, remove, and change the role of members in a Team or Enterprise organization. Member-added plugin administration Lets Claude for Government admins view and control member-added plugins from the Config page alongside the tool and connector cards. Model entitlements Lets Enterprise admins set which models and effort levels members can use, applied to the whole organization or to a single role. Organization instructions Lets admins define instructions of up to 3,000 characters that Claude applies to every conversation for every member and that take precedence over individual user instructions. Organization plugin marketplace Lets Team and Enterprise owners publish a curated plugin marketplace and control which plugins members can see and install. Organization skill management Lets admins provision skills and plugins, control user-created and shared skills and plugins, and scan third-party uploads for malicious content. Public project controls Lets Team and Enterprise owners stop members creating publicly shared projects while keeping internal sharing available. Role-based access (Claude) Lets admins assign role-based permissions to workspace and organization members. Skill and plugin security scanning Scans third-party skills and plugins for malicious content when an organization member uploads or edits them, as an Enterprise setting an admin turns on. Verified-domain connector restriction Stops Claude accounts outside the Enterprise organization from connecting services such as Gmail and Slack with an address on the organization verified domains.
Ecosystem connectors 1Password for Claude Lets Claude request a login from 1Password during a browser task so the credential is filled on the page without Claude seeing the password or one-time code. Claude for Microsoft 365 / Outlook Embeds Claude inside Microsoft 365 and Outlook so users can invoke it from within those applications. Claude for Outlook Runs Claude as an Outlook add-in that works across the inbox and calendar to triage mail, draft replies, and track threads. Claude LTI for Canvas Installs Claude as an LTI 1.3 app in Canvas LMS, placed in course navigation and assignment editing, and enabled from Claude for Education organization settings. Connected app suggestions Brings up a relevant connected app on its own when the conversation calls for it, instead of requiring the user to name it. Connectors Lets users attach remote MCP servers and pre-built connectors, such as Slack and Google Workspace, so Claude can read and act on external data during chat. Custom connectors Lets a user register their own remote MCP server as a connector and authenticate it to their account. Customization directory Browses skills, connectors, and plugins in a single directory, including the ones an organization has shared with its members. Desktop extensions (MCPB) Packages local MCP servers as MCPB desktop extensions that run in the Claude desktop app, with admin and MDM controls to enable, block, or upload custom extensions. GitHub connector Connects selected GitHub repositories so Claude can read repository content as context for development tasks. Google Workspace connectors Connects Gmail, Google Calendar, and Google Drive so Claude can search mail, manage calendar events, and read and save files. Interactive connectors Renders connector output as interactive apps inside a mobile conversation, including live charts, sketched diagrams, and shareable assets. MCP server submission Lets developers submit remote and local MCP servers for listing in the Claude connectors directory. Microsoft 365 connector Connects Claude to Microsoft 365 to search SharePoint, OneDrive, Outlook, and Teams, with admin-gated write tools for email, calendar, files, and Teams messaging. Plugins in Claude Installs plugins that bundle skills, connectors, and sub-agents into one package usable in chat, Claude Desktop, and Cowork, and shares a plugin a member built with named colleagues or, on Enterprise, a group. Skills in Claude Lets a user enable, browse, and create skills, folders of instructions, scripts, and resources that Claude loads on demand for specialized tasks. Tool access modes Controls how a large connector library is loaded into a conversation so connector definitions do not crowd out working context.
Plans & billing Central billing and administration Holds an organization's billing email recipients, payment method, invoice collection method, and monthly or annual billing cycle on one subscription that covers every member. Claude Enterprise Admin API Gives Enterprise organizations endpoints to read and set per-member usage-credit spend limits and to review and act on members limit-increase requests. Claude for Education Provides a university-sponsored Claude offering with institution-wide administration, sponsored seats for students and faculty, and campus deployment. Claude for Teachers Gives verified US K-12 educators free access to Pro features with teaching skills and standards-mapped curriculum under K-12 specific terms. Extra usage credits Lets an organization turn on usage-credit billing so members can keep using Claude and Claude Code at standard API rates after reaching the usage included in their plan. Gift subscriptions Buys a Pro or Max subscription as a gift and redeems a gift code onto an account. Organization spend limits Lets Enterprise admins set usage-credit spend caps in claude.ai settings at the organization, seat-tier, group, and member level, and resolve members' in-product requests for a higher cap. Seat management Lets admins purchase standard and premium seat tiers on a Team or Enterprise subscription, mix tiers within one organization, change or remove a member's tier, and map tiers from SSO groups. Self-serve Enterprise plans Lets an organization buy an annually billed Enterprise plan on the website without a sales conversation. Subscription management Lets a subscriber switch a Pro, Max, or Team plan between monthly and annual billing, cancel it, and request a refund. Tax details and invoices Stores a billing address and tax or VAT ID on a subscription or Console organization and provides downloadable receipts and invoices. Team to Enterprise upgrade Upgrades a Team organization to Enterprise in place, preserving memberships, roles, conversations, and projects. Usage and length limits Meters each plan against five-hour and weekly usage limits with in-product warnings, and caps how long a single response can be. Usage bundles Pre-purchases discounted usage-credit bundles that draw from one balance across Claude, Claude Code, and Cowork. Usage credits on paid plans Lets Pro and Max subscribers keep working past their plan included usage by paying for additional usage at standard rates.
Conversational core Automatic model switching Moves a conversation to a different Claude model when a request is blocked on a restricted model, and lets the user control whether switching happens. Chat context compaction Summarizes earlier messages when a chat approaches its context limit so the conversation continues past the previous length limit. Claude chat Provides Claude's conversational chat interface across web, iOS, Android, and desktop, including voice input. Code execution (chat) Lets Claude create, edit, and execute code and files directly within a chat conversation. Conversation management Lets a user rename or delete an individual conversation and clear all conversations from chat history. Crisis helpline resources Surfaces location-matched crisis helpline resources in a dismissable banner when a detection system separate from the model identifies a conversation touching on distress, suicidal thoughts, or self-harm. Dictation Converts speech to text in the prompt composer on the Claude mobile apps and in the Excel, PowerPoint, Word, and Outlook add-ins. Extended context window in chat Gives paid plans a larger per-conversation context window, up to 1M tokens on current Opus and Sonnet models. Interface preferences Lets a user set the interface language and choose light, dark, or system colour mode on web and desktop. Memory (chat) Remembers context and preferences across separate conversations to personalize future chats. Model and effort level selection Lets Claude Desktop on 3P deployments choose among offered Claude models, with a default model and starting effort level, per-model effort caps, and 1M-context variants. Model, effort, and thinking settings Lets a user pick which Claude model a chat uses, how much effort it spends on each response, and whether extended thinking is on. User preferences Lets a user set standing conversation preferences that shape how Claude responds across their chats. Voice mode Lets a user hold a spoken conversation with Claude on mobile, desktop, and web, with Claude speaking its replies and using connected tools.
Browser & device surfaces Claude deep links Handles claude:// URLs on desktop and mobile to open a chat, a Cowork session, or a Claude Code session with prefilled parameters. Claude Desktop Provides the Claude desktop application for macOS, Windows, and Linux, hosting chat, Cowork, and Claude Code sessions on the user machine. Claude Desktop enterprise deployment Lets admins install and update Claude Desktop centrally with pkg, dmg, or MSIX packages through Intune, SCCM, Group Policy, or PowerShell, and control the app through system policies. Claude mobile apps Provides the Claude apps for iOS and Android, distributed through the App Store and Play Store in supported regions. Computer use (desktop) Lets Claude drive the user's computer in Cowork and Claude Code by clicking, typing, and navigating desktop apps, gated by per-application permission prompts and an app blocklist. Health and fitness data on mobile Reads a user's health and fitness data on iOS and Android and returns activity, workout, and sleep analysis rendered as native charts. Mobile app integrations Connects Claude on iOS and Android to device apps so it can draft calendar events, find locations on maps, and manage reminders. Mobile widgets and shortcuts Provides home-screen widgets on iOS and Android plus the Ask Claude App Intent, Siri and Spotlight entry, and lock-screen and Control Center controls on iOS. Quick entry Opens Claude from anywhere on macOS to start a chat, capture a screenshot, share an application window, or dictate without leaving the current task.
Creative & generated content Artifact publishing Publishes an artifact to a public link or shares it inside the organization, and lets other users copy a published artifact into their own chat. Artifacts Lets Claude generate and display standalone rich content, such as code, documents, and visualizations, in a dedicated panel outside the chat flow. Claude Design in conversations and Artifacts Creates and edits designs with Claude Design inside conversations, the Artifacts tab, Claude Code, and mobile, with on-canvas editing and design-system import. Claude Docs Creates documents with Claude Docs in conversations and the Artifacts tab, gated by its own organization setting. Claude Slides Creates on-brand slide decks from notes, reports, or an existing conversation, with export to PowerPoint or PDF. Custom visuals in chat Generates charts, diagrams, and other visualizations inline in Claude's chat responses.
Usage insights Analytics chat Answers plain-language questions about an organization Claude usage by running the query and returning a chart with a short summary. Claude Enterprise Analytics API Gives Enterprise organizations programmatic access to usage and engagement metrics, enabled and keyed only by the Primary Owner. In-product surveys Lets Enterprise admins create short in-product surveys targeted at chosen groups, shown as a card in Cowork or chat, with aggregated responses reviewed in the Admin console's Analytics section. Smart reports Generates AI-analyzed usage reports for Enterprise admins covering workstreams, deliverables, cost per session, frictions, and reusable skill opportunities across chat, Claude Code, and Cowork. Usage analytics (Claude) Gives admins a dashboard of organization-wide engagement, adoption, and usage across Claude products.
Collaboration workspace Chat sharing Shares a snapshot of a conversation either as a public link anyone can view or as a private invite to specific people by email, and lets the owner update the snapshot or unshare it. Project knowledge retrieval Automatically switches large projects to retrieval-augmented mode so Claude keeps context over growing project content on paid plans. Project visibility and sharing Lets a Team or Enterprise user set a project as private or organization-visible and share it with specific groups. Projects Lets users organize related chats and documents into projects and share them with teammates for collaborative review.
Data portability File uploads Uploads PDF, DOCX, CSV, XLSX, and image files into a conversation for Claude to read and work with. Memory import and export Imports memory from another AI provider into Claude and exports Claude memory for backup or migration. Organization data export Lets a Team or Enterprise primary owner export the organization conversation and user data. Personal data export Lets a Free, Pro, or Max user export their conversations and account data from privacy settings.
Knowledge & retrieval Chat search Lets users ask Claude to search their past conversations and carry the retrieved context into the current chat. Enterprise search Searches across an organization's connected enterprise content in addition to the open web. Research Performs deep, multi-step web research and synthesizes the findings into a report. Web search (chat) Lets Claude search the web from within a chat conversation.
Personal reflection Break reminders and quiet hours Lets a user set optional break reminders and quiet hours from Claude's Time and focus settings. Monthly recap Shows a user the topics they spent time on, their most active day and peak hour, and observations about how they work with Claude over the past month.
Agentic work Agent approval modes Sets whether Claude asks permission before each agent action, reviews each action itself and blocks unsafe ones, or skips approval entirely, across Cowork, Claude in Chrome, connectors, and plugins.
Anthropic's agentic coding CLI that runs in a developer's terminal to read, edit, and run code in their own repositories.
Software capabilities CLI core Accessibility settings Screen-magnifier cursor tracking, a reduced-motion setting, and colorblind-friendly themes, usable independently of screen reader mode. Advisor (second-model consultation) Lets the main model consult a second, typically stronger model at key decision points — before committing to an approach, on a recurring error, before declaring done — applying its guidance without switching the whole session's model. Agent view keybindings Lets the agent view shortcuts Ctrl+S, Ctrl+T and Ctrl+G be rebound or unbound through keybindings.json via the agents:switchView, agents:togglePin and chat:externalEditor actions. Agentic coding loop Runs an autonomous gather-context/act/verify loop using built-in file, search, execution, and code-navigation tools (Read, Edit, Write, Bash/PowerShell, Grep, Glob, LSP) that Claude chains together to complete a coding task, interruptible at any point. Artifact publish validation Validates that an artifact's source file decodes as UTF-8 (or UTF-16 with BOM) before upload and refuses the publish with a positioned error when it does not. Artifact tab icons Lets Claude pick a browser-tab icon to match each artifact it publishes. Artifacts Publishes session output as a private, updatable interactive web page hosted on claude.ai, shareable privately, org-wide, or publicly, with admin controls over availability and retention. Automated security reviews Scans code for vulnerability patterns such as SQL injection, cross-site scripting, authentication flaws, and dependency issues, on demand through /security-review or on every pull request through a GitHub Action. Background monitor tool Watches a long-running command's output or a WebSocket feed in the background and interjects when a matching event arrives. Background session control Manages background sessions from the shell: resuming a session in place with --bg, deleting refused sessions with --discard-unpushed or --force-remove-worktree, and reading session state through claude agents --json. Bash edit diff Shows a diff of the files a Bash command changed in the Bash tool result when the Bash tool handles file edits, controlled by the bashEditDiffEnabled setting. Bundled skills Ships default prompt-driven workflows such as `/simplify`, `/batch`, `/debug`, `/loop`, and `/deep-research` that users invoke directly or Claude triggers automatically. Checkpointing (rewind) Automatically snapshots files on every prompt so a user can restore code, conversation, or both to an earlier point via `/rewind` or double-Esc. Claude Agent SDK Lets developers build their own agents on the Claude Code harness, metered against the signed-in plan usage limits. Claude API eval tooling The /claude-api skill builds eval sets for Claude-powered apps (build-eval) and iteratively improves an app against an existing eval (hillclimb). Claude Code authentication Signs Claude Code in with a Claude subscription or an API key, with an ANTHROPIC_API_KEY environment variable taking priority over the authenticated subscription. Claude Code CLI Runs Claude as an agentic coding assistant in a developer's terminal, reading, editing, and executing code in their own repository. Claude Code on the web Runs Claude Code tasks against a GitHub repository in an isolated remote virtual machine, several in parallel, and opens a pull request with the finished work. Claude Code plugins Lets users install and invoke third-party or organization-authored plugins inside Claude Code and Claude Cowork. Claude in Xcode Provides Claude-powered coding assistance inside Xcode for Pro and Max subscribers and premium Team and Enterprise seats. Claude-drafted feedback Drafts a feedback report with the SendFeedback tool when tools fail or requests can't be helped, shown as a reviewable card with a /feedback queue; nothing is sent until the user sends it. CLAUDE.md project memory Reads a CLAUDE.md file at the start of every session in a directory so project conventions and context carry into each run. Code intelligence (LSP) Language-server-backed code navigation — go-to-definition, find references, symbol search — and automatic post-edit type-error reporting, via installable per-language plugins. Code Review Runs a fleet of agents over a GitHub pull request and posts severity-ranked findings as inline comments, triggered on PR creation, on every push, or by an @claude review comment. Compaction context restoration Restores working context after conversation compaction: re-reads up to five recently modified files, re-injects invoked skill bodies within token caps, and keeps background commands and subagents running while reminding Claude which are still active so it does not start duplicates. Context window management Automatically compacts the conversation as the context window fills, with a configurable auto-compact threshold, a manual `/compact` with focus instructions, and `/context` to visualize what is loaded. Copy on select Copies selected text to the clipboard automatically when the mouse selection finishes in fullscreen rendering or agent view, with a global-config toggle to disable it. Cursor configuration import The /import command brings Cursor configuration — instruction files, MCP servers, commands, subagents, and skills — into Claude Code. Custom keybindings Lets users remap keyboard shortcuts, including chord sequences, across every UI context via a JSON config file. Customizable status line Renders a user-authored shell script as a persistent status bar showing model, cost, context usage, git branch, or any custom data. Dynamic workflows Runs large engineering tasks such as migrations, audits, and codebase-wide bug hunts end to end in a single session, with access controlled per custom role, through managed settings, or organization-wide. Fast mode Research-preview toggle for up to 2.5x faster Opus responses at higher per-token cost, billed against usage credits. Feedback drafts Drafts a session feedback report for the user to review and send from /feedback. Fullscreen rendering Flicker-free alt-screen terminal rendering with mouse support (click, drag-select, scroll) and constant memory usage in long conversations. Gerrit change lookup Looks up the open Gerrit change matching HEAD's Change-Id on googlesource.com checkouts when a Claude Desktop Code tab session starts or resumes. Git worktree isolation Isolates a session or subagent in its own git worktree and branch so parallel work never collides, with automatic cleanup and PR-based worktree creation. GitHub Actions integration Runs Claude inside a customer's own GitHub Actions workflows for automation beyond the packaged security and code reviews. Glob and Grep platform defaults Leaves Glob and Grep out of the default tool set on macOS, Linux, and WSL, where searches run through Bash with embedded bfs and ugrep, and restores the tools via CLI flags, Bash removal, or subagent tool lists. Goal-directed auto-continue (/goal) Sets a natural-language completion condition that a fast model judges after every turn, keeping Claude working automatically without re-prompting until the goal is met. IDE integrations Runs Claude Code inside VS Code, Cursor and other VS Code forks under the same account and usage limits as the terminal. Image and visual input Accepts pasted, drag-dropped, or referenced screenshots and mockups in prompts for visual analysis and design-to-code comparison. Immediate settings reload on /cd Applies the new directory's project settings, hooks, .mcp.json servers, skills, and subagents immediately after /cd instead of on the next --resume. Import config from other coding agents Migrates instruction files, MCP servers, commands, subagents, and skills from OpenAI Codex or Google Gemini CLI into Claude Code via `/import`. Interactive clarifying questions Lets Claude pose structured multiple-choice questions mid-task to gather requirements, with a configurable auto-continue timeout. Issue reference links Renders owner/repo#123 issue references as clickable hyperlinks to the issue on the repository's detected git host, with a FORCE_HYPERLINK override. JetBrains IDE plugin Plugin for IntelliJ, PyCharm, WebStorm, and other JetBrains IDEs that shells into an installed `claude` CLI in the IDE's terminal, adding IDE-native diff viewing, selection and tab context sharing, and diagnostic sharing. Large-codebase (monorepo) configuration Scopes Claude to the working subsystem of a monorepo via per-directory CLAUDE.md and skills, excluded-path rules, and denial of generated or vendored paths. Live diff panel Opens /diff as a live panel beside the conversation in fullscreen rendering, refreshing on edits and shell commands and allowing line selection for the next prompt. Mid-prompt command completion Completes slash commands typed partway through a prompt, matching command and plugin skill names after a mid-message '/' and inserting the full command with Tab. Mid-stream response continuation Continues a response cut off mid-stream by a server error, dropped connection, or stall in non-interactive sessions when the partial response has text and no tool calls. Model selection Sets which Claude model a session uses through the /model command, the --model flag, or the ANTHROPIC_MODEL environment variable. Non-interactive (headless) mode Runs a single prompt non-interactively via `claude -p` with text, JSON, or streaming-JSON output and JSON-schema-validated structured output, for scripting and CI use. Output styles Persona and format presets (Default, Proactive, Explanatory, Learning, or custom) that rewrite the system prompt to change Claude's tone, autonomy, and response format. Output styles Changes how Claude formats responses with selectable output styles, applied via /config, /output-style, or the styles interface, and supported mid-session. Screen reader mode Renders the CLI as plain labeled sequential text with numbered menus and terminal bell cues instead of spinners and in-place redraws, for screen reader and braille users. Session management Persists, resumes, names, forks, and branches conversations tied to a working directory, with a searchable session picker and PR-linked lookup. Setup diagnostics (/doctor) Diagnoses install health, invalid settings, unused extensions, slow hooks, and CLAUDE.md bloat, proposing fixes to apply on confirmation. Side questions (/btw) Answers an ephemeral, tool-free question about the current session in a dismissible overlay that never enters conversation history, forkable into a new session. Slash commands In-session `/command` system for session control, git and PR workflows, and configuration, extensible with user-authored skills. Task list tracking Maintains a Claude-managed to-do checklist for multi-step work, visible and toggleable in the UI and persisting across context compaction. Terminal configuration Configures terminal behavior — newline binding, desktop notifications, tmux passthrough, color themes, and Vim-style prompt editing. Voice dictation Transcribes spoken prompts into the CLI via hold-to-record or tap-to-record, tuned for coding vocabulary and multiple languages. Web fetch tool Built-in WebFetch tool that retrieves a URL, converts HTML to Markdown, and runs an extraction prompt against the content with a small fast model, prompting for approval the first time it reaches a new domain outside a preapproved documentation set. Web search tool Built-in WebSearch tool that queries Anthropic's web search backend for result titles and URLs, scoped with allowed- or blocked-domain lists, for Claude to follow up on with WebFetch. WebFetch deadline Fails WebFetch with a deadline error after five minutes instead of hanging; CLAUDE_CODE_WEBFETCH_DEADLINE_MS changes or removes the limit.
Extension system claude.ai-hosted plugin marketplaces Lists and adds plugin marketplaces hosted on claude.ai, such as an organization's plugin library, from the /plugin Marketplaces tab or via `claude plugin marketplace add --claudeai`; requires Claude Code v2.1.273 or later. Contextual plugin suggestions Lets marketplace operators declare contextual signals (working directory, CLI commands, files read, manifest dependencies) so Claude Code suggests a matching plugin to users, gated by an org-managed allowlist. Headless plugin reload The /reload-plugins command reloads active plugins without restart in non-interactive mode, the Agent SDK, and the desktop app. Hooks Registers shell, HTTP, MCP-tool, or prompt handlers that fire deterministically at fixed lifecycle events — tool use, session start/end, compaction, subagent start/stop — for enforced guardrails distinct from advisory instructions. Managed MCP configuration Lets admins deploy a fixed, exclusive MCP server set or enforce allowlist/denylist policies restricting which MCP servers users can add or connect to org-wide. MCP client Connects to external MCP servers over HTTP/SSE/stdio/WebSocket for tools, resources, and prompts, with OAuth auth, scoped installs, and a graphical Connectors setup flow in Desktop; can also run Claude Code itself as an MCP server. Plugin CLI JSON output Prints plugin install, uninstall, update, enable, and disable results as one JSON object on the last line of stdout with --json. Plugin evals Lets plugin authors write eval cases for a Claude Code plugin, run them with `claude plugin eval`, grade the results against a no-plugin baseline, and gate CI on the score. Plugin install --accept-command Lets non-interactive plugin installs accept a marketplace-declared command by passing the sha256 a previous -- run reported. Plugin install deep link Provides a vscode:// install-plugin deep link that opens the VS Code extension's Manage plugins dialog on one plugin's scope choice, prompting for a missing marketplace first. Plugin install hints Lets a third-party CLI signal Claude Code (via a stderr marker) to prompt a one-time install of that tool's companion plugin from the official marketplace. Plugin install with marketplace source Installs a plugin from a marketplace that has not been added yet by naming the marketplace source with `--marketplace` on `/plugin install`, resolving the source and confirming before adding; requires Claude Code v2.1.275 or later. Plugin local dependencies Resolves plugin dependencies from local --plugin-dir copies, including loading every sibling plugin when a parent folder is passed. Plugin marketplaces Catalogs for distributing and installing plugins from GitHub, git URLs, npm, or zip archives, with versioning, release channels, auto-update, dependency resolution, and org allow/block controls. PreModelSwitch hook gating Lets a PreModelSwitch hook approve, deny, or require confirmation for model switches, including switches requested by Agent SDK hosts and Remote Control clients. Session auto-archive Automatically moves inactive VS Code extension sessions to Archived sessions after a configurable number of days, controlled by the archiveInactiveSessions setting. Skill usage reporting Runs /skill-doctor to report each skill's context cost and invocation frequency, flagging never-invoked skills and recently unused plugins so users can decide which to turn off. Skills Reusable SKILL.md instruction files, auto-loaded when relevant or invoked with `/name`, supporting arguments, dynamic shell-command context injection, and tool pre-approval. VS Code MCP management dialog Lets users add, remove, enable, disable, reconnect, and OAuth-manage MCP servers from the /mcp dialog inside the VS Code extension without the terminal.
Enterprise deployment Bring-your-own-cloud model hosting Runs Claude Code against Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, or Claude Platform on AWS instead of the direct Anthropic API, each with its own auth setup and region configuration. Claude apps gateway Anthropic's self-hosted gateway (`claude gateway`) that signs developers in via corporate OIDC SSO, enforces per-group model access and managed settings, relays telemetry, and routes to Bedrock, Vertex, Foundry, or the Anthropic API with failover. Cloud gateway policy enforcement Enforces administrator managed-settings requirements that sessions sign in through the Claude apps gateway, rejecting locally configured first-party credentials. Corporate launcher Routes every process Claude Code starts — the background supervisor, agent-view sessions, Remote Control workers — through an org-mandated launcher binary for sandboxing, network control, or credential injection. Desktop admin controls Admin-console and managed-settings controls specific to Desktop — enabling or disabling the Code tab, web sessions, Remote Control, and Bypass Permissions — plus MDM and Group Policy deployment. Enterprise network configuration Supports corporate HTTPS proxies, custom or system CA trust stores, and mTLS client certificates, with a documented outbound network allowlist for running behind a firewall. Gateway on owned public address space Lets administrators declare organization-owned public IPv4 blocks in the gatewayInternalNetworks managed setting so /login accepts a Claude apps gateway addressed inside them when the client connects from the same block. Gateway pricing sync Delivers Claude apps gateway pricing rates to signed-in Claude Code clients through managed settings so /cost and telemetry match the gateway spend meter. Gateway public-client warning Warns at boot and at runtime, emitting an access.public_client audit event with the client IP, when a gateway with an empty access_control allow list receives a request from a public address. GitHub Enterprise Server support Connects Claude Code to a self-managed GitHub Enterprise Server instance via a one-time GitHub App install, enabling web sessions, code review, and security scanning without per-repository setup. LLM gateway support Routes Claude Code's API traffic through a customer-operated LLM gateway for centralized credentials, usage tracking, cost controls, and audit logging. Managed MCP servers in gateway policies Gateway managed policies can carry managedMcpServers in the cli block, providing remote MCP servers (http and sse) to every matching developer alongside the servers they add themselves. OIDC scope on refresh Lets gateway admins set oidc.scope_on_refresh so the gateway re-sends the scope list on refresh-token exchange for IdPs like Okta that return an id_token on refresh only when openid is requested again. Per-user 429 handling On upstreams with forward_user_identity, the gateway treats a 429 returned for a request carrying the developer's email as a per-user denial and returns it instead of failing over to the next upstream. Runner drain marker file The self-hosted runner accepts a drain marker file that a host writes to announce a graceful drain, so the runner reports its exit as a host drain rather than a plain shutdown signal. Runner host config snapshot modes The self-hosted runner keeps its startup snapshot of the host config directory on disk with per-file digest verification or in memory under a 64 MiB cap, controlled by --host-config-snapshot. Self-hosted cloud environments Runs cloud and web sessions on runner infrastructure the customer operates instead of Anthropic-managed VMs, with autoscaling orchestration, lifecycle hooks, and a signed session token services can verify. Self-hosted session state cleanup Optionally deletes each session's per-session directories under the self-hosted runner's base directory when the session ends, via claude self-hosted-runner --remove-session-state.
Agent SDK Agent SDK cost tracking Exposes per-step, per-model, and cumulative token and cost accounting on the SDK message stream, including cache breakdown and subagent spend rolled up separately. Agent SDK custom tools and MCP connectivity Extends an SDK agent with new capabilities via in-process function tools or external MCP servers over stdio/HTTP/SSE, sharing the same naming and permission mechanics. Agent SDK dynamic tool search Scales an SDK agent to thousands of registered tools by withholding schemas from context and having the agent search for and load only the tools it needs on demand. Agent SDK file checkpointing Tracks file changes during an SDK session and lets the host application programmatically rewind files to any prior checkpoint, independent of git. Agent SDK message stream Streams Agent SDK responses as one AssistantMessage per non-empty content block (text or tool call), with shared message IDs, and TypeScript StreamEvents carrying user_message_uuid for turn attribution. Agent SDK OpenTelemetry observability Exports SDK agent traces, metrics, and structured log events over OpenTelemetry to any collector, including trace-context propagation into the host application's own spans. Agent SDK real-time output streaming Emits raw partial output events as text and tool-call inputs are generated, for building live-updating UIs instead of waiting for complete messages. Agent SDK self-hosted deployment Documents and supports running the SDK on a customer's own infrastructure — subprocess lifecycle patterns, container or VM isolation options, credential-proxy patterns, and a programmatic sandbox API for network-restricted execution. Agent SDK session management Lets a host application resume or fork a conversation across process restarts, mirror transcripts to a self-managed storage backend, and run a persistent interactive session accepting queued messages and mid-task interrupts. Agent SDK structured outputs Constrains agent output to a JSON Schema, validating and auto-retrying until the result matches, returned as typed data instead of free text. SDK MCP server provenance Exposes the MCP server serving a tool and where its definition came from to SDK hooks, permission callbacks, and system init messages for trust decisions. SDK mid-session configuration Lets Agent SDK sessions change model, permission mode, and a set of settings keys while running, via setModel, setPermissionMode, applyFlagSettings, and updateSettings. SDK readFile scope Restricts the SDK control readFile() call to regular files inside the session's working directories plus a few of Claude Code's own session files, resolving with null for anything else, independent of Read tool allow rules. SDK startup failure reason Exposes a structured startup_failure_reason in stream- results, including worktree_unverified and worktree_resume_refused values, so applications can branch instead of matching error text. SDK system prompt snapshot control Lets Agent SDK hosts control system prompt reuse across requests: pass a custom prompt via the { type: 'custom' } form and set snapshot to reuse the prompt recorded on the session's first request or false to rebuild it every request. Subagent CLAUDE.md omission Lets a TypeScript SDK subagent run without the user, project, and local CLAUDE.md files while managed policy files still load. Task-tracking tool availability Provides the task-tracking tools (TodoWrite or the four Task tools) by default only on Claude 3.x, Opus 4 through 4.7, Sonnet 4 through 4.6, and Haiku 4.5, with opt-in via allowedTools or the tools option on other models.
Multi-agent orchestration Agent structured output Lets workflow agent() calls return JSON matching a caller-supplied schema, with pre-start contradiction checks and configurable validation retries. Agent teams Experimental mode where a lead session spawns teammates that each run as full independent sessions, self-coordinating via a shared task list and direct peer-to-peer messaging. Agent view (background sessions) Dashboard (`claude agents`) to dispatch, monitor, peek at, reply to, and attach or detach from many independent local or background sessions, including PR-status tracking. Configurable workflow concurrent-agent cap Lets users configure the workflow runtime's concurrent-agent cap from 1 to 256 via the CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS environment variable. Cross-session messaging Lets a session discover and message the user's other independent local sessions, or reply to sessions on other machines via Remote Control, with configurable inbound accept/hold/refuse rules. Projects (new version, beta) Runs a beta version of projects in Claude Code where one conversation breaks work into parallel cloud threads that keep running and share the project's files, instructions, and memory. Projects coordination Lets a user give Claude a body of related work in one conversation and have it coordinate parallel cloud sessions that share repositories, instructions, and memory. Subagent partial output resume Marks a subagent's output as partial at its maxTurns limit, with a hint that Claude can continue it via SendMessage. SubagentHandback tool Delivers a subagent's final report through a dedicated tool so the auto mode classifier can review it before the report reaches the parent conversation. Subagents Delegates a focused task to a specialized assistant with its own context window, system prompt, tool access, and model, returning only a summary to the parent conversation. Teammate revival on message Restores a stopped in-process agent teammate in the same session when Claude messages it, re-applying its definition only from trusted folders. VS Code agent map Opens an agent map from the VS Code extension's agent count to view a subagent's read-only transcript or stop it. Workflow size guidelines Sizes dynamic workflows with a size guideline (unrestricted, small, medium, large) configurable in /config or settings files. Workflow usage-limit pause Pauses a workflow run when an agent hits the claude.ai usage limit, resuming the waiting agents automatically after the reset instead of failing them.
Permission configuration Auto mode A permission mode where a separate classifier model reviews and auto-approves routine actions in the background, blocking scope-escalating or untrusted-destination actions without prompting the user. Auto mode classifier Auto-approves tool calls in auto mode subject to a background safety classifier that blocks data exfiltration, including paste-service links, and names the matched rule on denial. Bash permission rules Matches Bash permission rules against command text with wildcards, documenting compound-command, wrapper, symlink, and redirection matching and what a rule does not match. Dispatch permission mode resolution Resolves the permission mode of a session dispatched from agent view from the target directory's permissions.defaultMode first, falling back to the dispatching session's mode, with permissiveness guards. Fine-grained permission rules Allow/ask/deny rules scoped by tool and argument pattern — bash commands, file paths, domains, MCP tools, subagents — configurable per user, project, or org via managed settings. Permission modes Session-wide modes — Manual, Accept Edits, Plan, Auto, Bypass — controlling how often Claude pauses for approval, switchable mid-session or set as an org default. Permission modes Controls how tool calls are approved across manual, acceptEdits, plan, auto, dontAsk, and bypassPermissions modes, with dontAsk still running reads while denying anything that would prompt. Permission rule negation patterns Supports gitignore-style negation patterns in deny and ask path rules, carving matched paths out of earlier path rules within the same settings source. SDK permission modes Governs Agent SDK tool approval through permission modes and allow/deny rules, with dontAsk and default modes now running calls that need no approval in default mode, such as read-only Bash commands and in-workdir file reads. Server-managed settings Delivers organization policy — permission rules, sandboxing, MCP and plugin restrictions, model allowlists, hooks, login restrictions, version floors — from the admin console to clients, refreshed hourly and taking precedence over local settings. Settings hierarchy Layered configuration system — managed, user, project, and local scopes across settings.json files, MDM, or registry — with defined precedence and live reload for most keys. Subagent permission modes Runs each subagent in the parent session's permission mode unless its AgentDefinition sets permissionMode and the parent is in default, dontAsk, or plan mode; a definition can never apply bypassPermissions (v2.1.267+). VS Code hooks and permissions editors Adds or removes hooks and permission rules across user, project, and local settings from the VS Code extension's Customize menu. WebFetch artifact gating Applies WebFetch domain permission rules to reads of published artifacts, with bare WebFetch rules no longer blocking artifact reads.
Usage analytics /usage Loops breakdown Adds a Loops breakdown to /usage on Pro, Max, Team, and Enterprise plans with run count, total tokens, tokens per run, and last run for the heaviest /loop and scheduled tasks. Claude Code Analytics API Gives organization admins programmatic access to daily Claude Code usage and developer-productivity metrics, including sessions, lines changed, commits, tool acceptance, and cost. Claude Code analytics dashboard Shows plan owners and Console users lines of code accepted, suggestion accept rate, activity trends, top commands, and per-member usage, alongside a Value tab estimating productivity lift and cost per commit. Contribution metrics Connects a GitHub organization to compare pull requests merged and lines committed with and without Claude Code, at both organization and user level. Direct telemetry export to collector A managed policy can set OTEL_EXPORTER_OTLP_ENDPOINT to a collector's https URL so sessions signed in through /login export telemetry straight to that collector instead of through the gateway relay, subject to client-side checks. Gateway telemetry relay The gateway relays OTLP/HTTP telemetry to configured destinations with per-signal exporter selectors, stamps Desktop and Cowork exports with user.email and user.groups, and pauses a failing destination for 30-second stretches after five consecutive failed deliveries. OTel repository attributes Tags OpenTelemetry metrics and events with vcs.* repository identity attributes derived from the session's origin remote, gated by OTEL_METRICS_INCLUDE_REPOSITORY. OTel span diagnostics Emits bounded, privacy-safe span attributes for tool names, bash command classes, error classes, trace parenting, and request timing on OpenTelemetry spans. Prompt cache miss diagnostics Reports the likely cause of the most recent prompt cache miss and per-cause session counts through the statusline JSON fields last_miss_cause and miss_causes. Usage limit warnings Warns the user before a usage window runs out, naming the percentage used and the reset time.
Cost and usage governance Bedrock CountTokens spend accounting The gateway uses the Bedrock CountTokens API, at no charge, to count input tokens of abandoned requests so spend limits stay accurate, falling back to a one-token request when the permission is absent. Claude apps gateway spend limits An admin API for the Claude apps gateway that caps developer, group, or organization spend by day, week, or month, enforced live on every request. Contracted-rate cost reporting Reports /usage, status line, and OpenTelemetry cost at contracted rates via the modelPricing managed setting instead of list price. Gateway pricing markup Allows a Claude apps gateway pricing multiplier above 1, up to 10, to meter requests at a markup such as an internal chargeback rate, optionally applying the markup to spend limits. OpenTelemetry usage export Exports Claude Code usage metrics, logs, and (beta) distributed traces via OpenTelemetry to a customer-operated observability backend, independent of Anthropic-hosted analytics. Prompt cache TTL settings Sets promptCacheTtl (e.g. 1h) for the main conversation's prompt cache and subagentPromptCacheTtl for subagents and other requests when using an API key or cloud provider. Session cost and attribution breakdown (/usage) Shows a per-session token-cost breakdown by model and, on paid plans, an attribution breakdown of recent usage by skill, subagent, plugin, or MCP server with flagged high-cost behaviors. Spend limit enforcement Blocks requests when a plan, individual, org, team shared budget, or Slack channel spend limit is reached, naming whose limit applied and how to raise it. Usage credits Lets a user, or an admin-approved Team/Enterprise member, purchase usage beyond the plan's session or weekly limit via `/usage-credits`.
Native apps Desktop app Native graphical app for macOS, Windows, and Linux with parallel sessions each auto-isolated in its own git worktree, drag-and-drop panes (chat, diff, browser, terminal, plan, tasks), and a permission-mode and model selector. Desktop CLI session resume Lets developers move a CLI session into the Desktop app with /desktop, or type /resume in Desktop to list CLI-started sessions searchable by title, folder, or branch, preview where each left off, and continue one with full context. Desktop pop-out panes Pops any Desktop pane such as the diff or terminal out into its own window and docks it back, for multi-screen layouts while Claude keeps working. Desktop sessions in WSL Runs a Desktop session inside a WSL 2 distribution so Claude Code, tools, and git execute with native Linux paths instead of over the Windows network filesystem. Desktop SSH remote sessions Runs a Desktop session on a remote machine over SSH while using the Desktop app as the interface, with admin-configurable pre-approved host allowlists. Dispatch (mobile-to-desktop task routing) Routes a task messaged from the mobile app into a new or existing Desktop Code session, with a push notification on completion or when approval is needed. iOS Simulator pane (Desktop) Desktop pane that streams and lets a user interact with Apple's iOS Simulator when Claude builds, runs, or tests an iOS app. Live app preview Starts or detects a dev server and opens it in an embedded Desktop browser pane, auto-verifying edits by screenshotting, inspecting the DOM, and interacting with the running app. Mobile app Mobile client into cloud sessions, a locally-running session via Remote Control, or the Desktop app via Dispatch, with QR-code pairing and push notifications.
Workspace governance Claude Code on subscription plans Lets Pro, Max, Team, and Enterprise subscribers run Claude Code against their subscription instead of Console API credits. Claude Code usage limits Meters Claude Code against the signed-in plan session and weekly usage limits and against per-session context limits. Claude Code workspace Keeps each Claude Code user's API traffic in a dedicated, separately rate-limited and spend-limited organization workspace created automatically at sign-in. claude.ai skill and plugin sync control Lets admins stop Claude Code from loading or syncing the skills and plugins developers enable on claude.ai, via the syncClaudeAiSkills and syncClaudeAiPlugins managed settings keys. GitLab-hosted marketplace sync Syncs plugin marketplaces hosted on gitlab.com or a self-managed GitLab instance through Organization settings, with per-host private-source rules. Managed effort cap Lets organizations cap the reasoning effort level for every model or per model, on every provider, through the `maxEffortLevel` managed setting. Managed MCP server control Lets organizations restrict which MCP servers users can add or connect to, deploy a fixed set, or provide remote servers to every user alongside their own, via managed settings keys and a deployed managed-mcp.json file. Managed MCP server provisioning Provides a set of remote MCP servers to every user through the managedMcpServers managed setting, alongside the servers users add themselves, without taking exclusive control of MCP. Managed settings approval dialog Shows a security approval dialog for organization server-managed settings and exits without applying them when the user declines.
Model configuration Advisor model Lets users set a stronger advisor model that Claude consults at decision points, via /advisor, and now also works in headless mode, the Agent SDK, the desktop app, and over Remote Control (v2.1.260+). Advisor non-interactive control The /advisor command accepts a model or off as an argument in non-interactive and Remote Control sessions and prints the current advisor as text when run bare. Automatic model fallback Automatically retries a request on a configured backup model when the primary model is overloaded, unavailable, or flagged by a safety classifier. CLI system prompt snapshot control Controls whether a resumed conversation reuses the system prompt recorded on its first request or rebuilds it on every request, via the --system-prompt-snapshot flag. Effort level and extended context controls Tunes reasoning depth via `/effort` (low through max) and opts into a 1M-token context window on supported models and plans. Extended thinking mode Toggles Claude's extended reasoning display and, on fixed-budget models, whether it thinks at all — via a session hotkey, a global `alwaysThinkingEnabled` default, or `MAX_THINKING_TOKENS` — with subagents inheriting the main session's setting. Organization model governance Lets Enterprise admins restrict which models members can run, set an organization default model, and cap effort levels per custom role, enforced server-side. Per-model effort levels Saves a separate effort level per model from /effort and the /model picker, with a session-only apply option.
Cloud and remote execution Cloud environment API credentials Stores an API key or token on an Anthropic-hosted cloud environment that the agent proxy attaches to requests for named hosts after they leave the session VM, so sessions call the API without the key entering the sandbox. Cloud environment configuration Configurable execution environment — network access level, environment variables, cached setup scripts — shared by every cloud session regardless of entry point: web, `--cloud`, Desktop, mobile, routines, or Claude Tag. Cloud session idle expiry Cloud sessions expire after inactivity, and a session waiting for the user to approve an MCP connector tool call or sign in to an MCP server counts as inactive and can expire during that wait. Cloud session sharing Shares a cloud session's live page via a private, team, or public link, with per-person viewer or editor roles on Team and Enterprise and admin control over public sharing. Local repository bundle upload Claude Code on the web bundles and uploads a local repository to a cloud session when there is no git remote or the Claude GitHub App is not installed on the github.com repository, forced with CCR_FORCE_BUNDLE=1, with push-back only when the GitHub connection has push access. Queued message recall Takes back a queued message in a web cloud session before Claude reads it, returning the text to the message box. Web setup token management Stores the GitHub CLI token encrypted against the claude.ai account for cloud-session GitHub access, with documented removal via connectors and revocation on GitHub.
Code security Claude Security plugin (local scan) Installable plugin that runs a multi-agent vulnerability scan of a local codebase or diff, writes a timestamped report with independently verified findings, and generates patches for the user to apply. Code review GitLab comments Posts /code-review findings as inline comments on GitHub pull requests and as a single note on GitLab merge requests via the glab CLI. Local diff review (/code-review) Terminal command that reviews the current branch's diff, a file, a PR, or a ref range for correctness bugs, running as a background subagent with adjustable effort levels and optional auto-fix. PR auto-fix and auto-merge Watches a pull request and automatically pushes fixes for CI failures or review comments, with a one-click auto-merge once checks pass on Desktop. Security guidance plugin (in-session review) Installable plugin that reviews Claude's own code changes for vulnerabilities at three depths — per-edit, end-of-turn, and on each commit or push — fixing what it finds in the same session. Ultrareview Deep cloud-based code review that launches a larger fleet of reviewer agents in a remote sandbox with independently verified findings, billed as usage credits with a limited free allotment on Pro and Max. Visual diff review and self-review File-by-file diff viewer with inline line comments across Desktop and the web, plus a self-review action where Claude critiques its own diff for bugs and security issues before the user does.
Sandboxed execution Dev container support Official Dev Container Feature that installs Claude Code into any Docker-based dev container with persisted auth, org-policy enforcement, and an egress-restricting firewall reference config. Per-command allowed domains in auto mode In auto mode with sandboxing on, Claude names the hosts a Bash, PowerShell, or Monitor command needs on the command itself, and the classifier reviews them with the command instead of prompting per connection. Restricted mode Starts Claude Code without command- and code-running tools via --restricted, removing WebFetch, confining file tools to working directories, loading only managed settings, and refusing bypassPermissions. Sandbox runtime package Standalone package that wraps the entire Claude Code process — not just Bash — in the same OS-level isolation, so built-in tools, MCP servers, and hooks share one configured filesystem and network boundary. Sandboxed Bash tool Per-command OS-level sandbox (macOS Seatbelt, Linux/WSL2 bubblewrap) restricting filesystem access and network domains for every Bash command and its children, configurable via `/sandbox` and enforceable org-wide. Sandboxed shell command coverage The sandboxed shell tool restricts the filesystem and network access of every Bash, PowerShell, or Monitor command and its child processes using operating system primitives. Worktree git-config safety refusals Refuses to create or sweep Claude-created worktrees when the repository git config defines filter drivers it cannot neutralize or settings it cannot switch off.
Developer tooling Artifact file downloads Lets artifacts offer viewers file downloads, such as CSV or PNG exports, through a claude.ai-enabled downloads capability declared at publish. Custom model picker entries Extends or replaces the /model picker's built-in list via the modelPicker setting with custom ordered, labeled entries including Bedrock and Google Cloud Agent Platform model IDs. Read shared artifacts Lets Claude read an artifact shared with the user by URL or /artifacts, summarizing it and saving the full source locally for exact retrieval. Team memory (Claude Code) Shares a Claude Code team memory store across the organization, with access governed by an ACL that grants or removes RBAC groups at per-group access levels. VS Code open-file attachment Automatically includes the file open in the VS Code editor as message context, with an attachOpenFile setting to restrict context to selected text only. VS Code prompt cache clock Shows a countdown of the conversation's prompt cache lifetime next to the context indicator so developers can anticipate slower, more expensive responses after expiry.
Team communication Channels (event integrations) MCP servers — Telegram, Discord, iMessage, custom webhook receivers — that push external chat messages or CI and monitoring events into a running session so Claude can react and reply. Claude Code in Slack Slack app that routes `@Claude` mentions to a Claude Code on-the-web session under the individual user's own account, with chat-vs-code routing and message actions. Claude Tag Slack integration that runs `@Claude` in team channels under a shared organizational identity with admin-configured access. Cross-session headless sender notices Sends headless `claude -p` senders delivery notices for cross-session messages as streamed informational system messages, telling the sender's Claude when a receiver holds, delivers, denies, or refuses a message; requires Claude Code v2.1.271 or later.
Automation and scheduling Desktop scheduled tasks Local recurring or one-off task scheduler in the Desktop app's Routines page, running on the user's machine with direct file access only while the app is open. Routines (cloud automation) Saved cloud-hosted configurations — prompt, repos, connectors — that run autonomously on a schedule, an authenticated API call, or GitHub repo events, without a local machine or open session. Scheduled tasks (/loop, cron) Runs a prompt on a recurring or one-shot basis within a session via `/loop` or Cron tools, persisting only while the session stays open or is resumed within 7 days.
Computer and browser automation Background computer use Runs computer use in the background on macOS in the Desktop app so Claude sees and acts in approved apps while the user keeps working; beta on Pro and Max. Claude in Chrome integration Connects Claude Code to the Claude in Chrome browser extension for live debugging, form filling, data extraction, and screenshot or GIF capture, sharing the user's logged-in browser session. Computer use Claude directly controls the screen — clicking, typing, scrolling, screenshotting — on macOS or Windows, gated by per-app tiered approval and a machine-wide single-session lock.
Data protection and compliance Console account sign-in Offers sign-in with a Console account under /login for members of Console organizations that don't allow API keys. Zero Data Retention For qualified Enterprise accounts, processes Claude Code prompts and responses without storing them after the response returns, with a defined list of disabled features and a BAA extension for healthcare customers.
CI/CD integrations GitLab CI/CD integration Runs Claude Code as a GitLab CI/CD job triggered by `@claude` mentions or pipeline events to create merge requests, fix bugs, and implement features.
Anthropic's API and developer platform for building with Claude, spanning the Messages API, built-in tools, Claude Managed Agents, and admin/compliance management APIs.
Software capabilities Admin & compliance Access Transparency Lets a customer retrieve tamper-proof log entries recording every read of their retained content by Anthropic personnel. Admin API Lets organization admins programmatically manage members, workspaces, roles, and API keys instead of using the Console UI. Compliance API Gives Claude Enterprise compliance teams programmatic access to the organization's activity feed, directory, and underlying chats, files, and projects, including a record of Anthropic personnel access. Console API keys Creates and manages API keys in the Console, scoped to a workspace and shown once at creation. Console feedback controls Lets a Console admin turn off members ability to send thumbs-up and thumbs-down feedback to Anthropic. Console member management Invites and removes members of a Console organization and lets people signing up with a business email find and join an existing one. Console roles and permissions Assigns Console members one of six roles that determine access to the Workbench, API keys, usage logs, and billing. Console workspaces Groups API keys into workspaces carrying their own spend limits, rate limits and roles. Content provenance marking Marks Claude-generated output with Anthropic's text watermark and embeds C2PA Content Credentials in supported image, video, and audio files produced by the code execution tool and retrieved through the Files API. Customer-managed encryption keys Encrypts Claude workspace data at rest with a key the customer provisions and controls in AWS KMS, Google Cloud KMS, or Azure Key Vault, retaining rotation and revocation. Data residency Controls where model inference runs, per request or as a workspace default, and where workspace data is stored at rest and endpoint processing happens. Real-time cyber safeguards Blocks prohibited and high-risk dual-use cybersecurity requests on Opus- and Sonnet-class models, with an application-based Cyber Verification Program that lifts dual-use blocks for an approved organization. Service accounts Creates, lists, updates and archives organization service accounts and manages their workspace memberships over the API. Workload Identity Federation Lets automated workloads mint short-lived Claude API credentials from an external identity provider, such as Kubernetes, cloud IAM, GitHub Actions, Okta, or SPIFFE, instead of static keys. Workspace data retention Turns on 30-day data retention for one Console Workspace inside a zero-data-retention organization, which Covered Models require, while the organization's other workspaces keep zero data retention.
Agent tooling Advisor tool Lets a faster executor model consult a higher-intelligence advisor model mid-generation, which reads the conversation and returns a plan or course correction. Bash tool Gives Claude an Anthropic-defined tool for requesting shell commands that the caller's application runs in a bash session it keeps alive across calls. Browser use tool Lets Claude navigate, read, and act on webpages in a browser the application runs, through a client toolset of 27 member tools plus optional JavaScript execution, file upload, and console and network reading, with download events rendered into Claude's context. Code execution tool Runs Python and bash code Claude writes in a sandboxed container so it can analyze data, run commands, generate files, and process uploaded files inside an API conversation. Computer use tool Gives Claude screenshot, mouse, and keyboard control of a desktop environment that the caller's application drives. Memory tool Gives Claude an Anthropic-defined file-operation tool for storing and retrieving information across conversations in a directory of memory files the caller hosts. Programmatic tool calling Lets Claude call the caller's tools from code running in the code execution container instead of taking a model round trip per tool invocation. Server-side tools Provides the shared mechanics for Anthropic-executed tools: the server_tool_use block, pause_turn continuation, turns that mix server and client tools, and domain filtering. Text editor tool Gives Claude an Anthropic-defined text editor tool with view, create, str_replace, and insert commands that the caller's application executes against its own files. Tool search tool Lets Claude search a tool catalog and load only the definitions it needs, instead of placing every tool definition in the context window up front. Tool use Lets callers define their own tools with JSON Schema and run the tool-use loop, where Claude emits tool_use blocks that the application executes and returns as tool_result blocks. Web fetch tool Lets Claude retrieve the full content of specified web pages and PDF documents from the Messages API, with dynamic filtering of fetched content. Web search tool Lets Claude search the web from the Messages API and answer with citations, with domain controls and code-based dynamic filtering of results before they enter context.
Performance & reliability Batch processing Processes large volumes of Messages API requests asynchronously at half the standard per-token price. Compaction Summarizes older conversation context server-side as a session approaches the context window limit, so long conversations continue without client-side summarization code. Context editing Clears selected content from conversation history at runtime under configurable strategies, including tool-result clearing and thinking-block clearing. Fast mode Serves supported Opus models at up to 2.5x higher output tokens per second at premium pricing. Long context windows Provides access to a 1M-token context window, gated per model and per deployment platform. Mid-conversation system messages and tool changes Lets callers append system-role messages and change tool availability partway through a conversation without invalidating the cached prefix ahead of them. Prompt caching Caches a prompt prefix so later requests resume from it, through automatic caching or explicit cache_control breakpoints with 5-minute or 1-hour TTLs. Rate limits API Lets admins query organization- and workspace-level API rate limits programmatically instead of through the Console UI. Refusal and fallback handling Detects classifier refusals and retries the request on a fallback model, server-side or through SDK middleware, without re-paying the full prompt-cache cost. Service tiers Routes Messages API requests through Standard, Priority (existing capacity commitments only), or Batch service tiers to balance availability, latency, and cost. Token counting Counts the tokens a request would consume before sending it, via /v1/messages/count_tokens.
Core API & extensibility Claude in Microsoft Foundry Serves Claude models through Microsoft Foundry so Azure customers call them under existing Azure agreements and commitments. Claude on Amazon Bedrock Serves Claude models through Amazon Bedrock so AWS customers call them under their own AWS account, regions, and billing. Claude on Google Cloud Vertex AI Serves Claude models through Google Cloud Vertex AI so Google Cloud customers call them under their own project, regions, and billing. Fine-grained tool streaming Streams a tool's input to the client as Claude generates it, without server-side JSON buffering or validation. MCP connector Connects the Messages API directly to remote MCP servers, with per-tool allow/deny configuration, without a separate MCP client. Messages API Provides direct, low-level access to prompt Claude models over a request/response HTTP API. Strict tool use Guarantees Claude's tool names and inputs match the tool's JSON Schema by constraining token sampling to schema-valid outputs. Structured outputs Constrains Claude's response to a caller-supplied JSON schema through output_config.format so downstream code can parse it without retries. User profile external user details Accepts platform-stated external user details — account status, country, entity type, hashed email and name, onboarding time, and reference ID — on user profiles under the user-profiles-2026-09-04 beta header.
Developer tooling Apple Foundation Models support Provides a Swift package that exposes Claude as a server-side language model conforming to Apple's Foundation Models framework, so iOS, macOS, visionOS, and watchOS apps can drive it through the same session API used for Apple's on-device model. Claude API CLI Provides the `ant` command-line tool for calling every Claude API resource from the terminal, with typed-flag or piped-YAML request bodies, response field extraction, and automatic pagination of list endpoints. CLI session connect Attaches the terminal to a Claude Managed Agents session with `ant beta:sessions connect`, following it live, sending messages, and approving or denying tool calls, or serving the Console session viewer locally with --web. Client SDKs Publishes official client libraries for Python, TypeScript, Java, Go, Ruby, PHP and C#. OpenAI SDK compatibility Lets developers call the Claude API using an official OpenAI SDK by changing only the base URL, API key, and model name. Prompt sharing Shares a Workbench prompt with workspace members for joint editing, keeping a revision history attributed to each contributor. Workbench Runs prompts against Claude models and API features interactively in the Console before moving them into code.
Managed Agents platform Claude Managed Agents Provides a pre-built, configurable agent harness that runs Claude autonomously in managed cloud infrastructure for long-running and asynchronous tasks. Dreams Reads an agent's memory store alongside past session transcripts and produces a reorganized store with duplicates merged and stale or contradicted entries replaced. Managed Agents auto permission policy Lets Managed Agents toolsets use an auto permission policy where the server judges each tool call from the tool, its input, and session content, allowing safe calls, denying high-risk ones, and asking on unresolved calls. Managed Agents memory Gives Managed Agents sessions persistent, versioned memory stores that carry user preferences, conventions, and domain context across sessions. Managed Agents multiagent orchestration Lets a coordinator agent delegate work to other agents within the same Managed Agents session, each in its own isolated thread. Managed Agents scheduled deployments Runs a Managed Agents agent automatically on a recurring cron schedule and tracks its run history.
Reasoning & output control Effort Lets callers set an effort parameter that controls how many tokens Claude spends on a response, trading response thoroughness against token use. Extended thinking Lets callers enable thinking with a fixed budget_tokens budget on Claude 4.5 and earlier models, deprecated on 4.6 and rejected by 4.7 and later in favor of adaptive thinking. Reasoning control Lets Claude work through a problem in visible thinking blocks before answering, with adaptive thinking setting how much it thinks per turn. Task budgets Lets callers give Claude an advisory token budget for an entire agentic loop, including thinking, tool calls, and output, so the model self-regulates and finishes gracefully as the budget is consumed.
Spend management Console usage and cost reporting Breaks down API usage and cost by model, workspace, and API key in the Console for Developer, Billing, and Admin roles. Prepaid API usage credits Bills API, Workbench, and Claude Code usage against prepaid usage credits that can be topped up manually or on auto-reload. Spend Limits API Lets admins set a spend limit on each Claude Enterprise member, see where a member's limit is inherited from, and approve or deny requests for a higher limit. Usage and Cost API Gives admins programmatic access to the organization's historical API usage and cost data, broken down by time bucket and dimension.
Managed Agents operations Managed Agents self-hosted sandboxes Lets an organization run a Managed Agents session's tool execution and filesystem on its own infrastructure instead of Anthropic's cloud sandboxes. Managed Agents vaults Stores MCP and environment-variable credentials for Managed Agents sessions so agents can authenticate to external services without exposing secret values. MCP tunnels Lets an organization securely expose MCP servers running in its private network to Claude over an outbound-only tunnel.
Multimodal content PDF support Lets Claude read and analyze the text and visual content of PDF documents. Vision Lets Claude understand and analyze images supplied as base64 data, URLs, or uploaded files.
Data handling Files API Lets callers upload a file once and reference it by ID across many Messages API requests instead of re-uploading its content.
Developer extensibility Agent Skills Lets Claude load modular, filesystem-based Skill packages of instructions and scripts on demand across claude.ai, Claude Code, and the API.
Grounded retrieval Citations Grounds Claude's answers in supplied source documents and returns the exact passages that support each claim.
Anthropic products and featuresanthropic.com