Anthropic Claude Code

MCP servers connect the agent to tools, databases, APIs, issue trackers and monitoring systems at local, project and user scopes, with OAuth 2.0 sign in for remote servers and managed allowlists, denylists and per tool ask or block settings for organizations, alongside built in file, search, shell and web tools.

Sourcecode.claude.com/docs/en/mcpread 2026-09-29

Subagents run in their own context windows with their own prompt, model, tools and permission mode, can run in parallel or in the background, and include built in Explore, Plan and general purpose agents; agent teams let a lead session spawn and supervise coordinated sessions that message each other, and /batch splits large changes into parallel background subagents after a plan is approved.

Sourcecode.claude.com/docs/en/sub-agents and docs/en/commandsread 2026-09-29

The agent grounds itself in the customer's repository by reading and searching it directly on each run, with project instructions from CLAUDE.md or AGENTS.md and path scoped rules, and MCP connections reach the customer's databases, issue trackers and internal APIs. Context is assembled per run, and no index, graph or embeddings layer that persists and stays queryable is documented.

Sourcecode.claude.com/docs/en/memory and docs/en/mcpread 2026-09-29

In Manual mode the agent starts read only and asks before it edits files, runs tests or executes commands that modify the system, with approve once or allow from then on, fail closed matching for unmatched commands and a working directory write boundary; auto mode sends actions to a separate classifier and organizations can turn it off, and managed settings enforce permissions above user scope.

Sourcecode.claude.com/docs/en/securityread 2026-09-29

The security page points to a SOC 2 Type 2 report and an ISO 27001 certificate in the Anthropic Trust Center, and documents a permission based architecture, sandboxed shell commands with filesystem and network isolation, command injection detection, trust verification for new codebases and MCP servers, keychain credential storage, managed settings for organization enforcement and a HackerOne disclosure program.

Sourcecode.claude.com/docs/en/securityread 2026-09-29

OpenTelemetry exports metrics, events and optional traces, with a tool decision event for every permission accept or reject and a tool result event for every completed tool call, and OTEL_LOG_TOOL_DETAILS adds commands, MCP server and tool names and inputs; managed settings can lock the telemetry destination, and all operations in cloud sessions are logged for audit.

Sourcecode.claude.com/docs/en/monitoring-usage and docs/en/securityread 2026-09-29

Auto memory, on by default, is written by Claude itself as it works (user, feedback, project and reference notes), scoped per repository on the user's machine, loaded into every session through a MEMORY.md index, kept outside the transcript retention sweep until the user or Claude edits or deletes it, and browsable and editable through /memory. CLAUDE.md files are standing instructions the user writes, separate from this memory.

Sourcecode.claude.com/docs/en/memoryread 2026-09-29

Runs on the developer's own machine by default, with code execution and file access staying local even in Remote Control sessions; cloud sessions run in isolated Anthropic managed VMs with configurable network access, or an organization routes them to self hosted environments on its own infrastructure, and dev containers add isolation. The limit: the security docs state data is encrypted in transit but not at rest, and Zero Data Retention is an enterprise arrangement rather than a default.

Sourcecode.claude.com/docs/en/securityread 2026-09-29

Plugins package skills, subagents, hooks and MCP servers as one installable unit, and Anthropic's official marketplace is added on first run and browsable from the /plugin Discover tab, with named plugins such as code-review, commit-commands, security guidance and skill-creator that each do their own job; bundled skills (/code-review, /verify, /run, /batch, /simplify) and built in subagents ship with the product.

Sourcecode.claude.com/docs/en/plugins, docs/en/commands and docs/en/github-actionsread 2026-09-29

The Claude Code GitHub Action runs in automation mode on any GitHub event, such as a pull request being opened or a cron schedule, without anyone mentioning it, and Code Review reviews every pull request automatically; beyond that the agent answers @claude mentions and runs from the terminal, desktop, IDEs, the web and headless scripts.

Sourcecode.claude.com/docs/en/github-actionsread 2026-09-29

Every model Claude Code runs is Anthropic's: the customer picks among Claude models per session or subagent and picks where they are served (the Claude API, Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry), but not the model maker, and the gateway documentation states that Anthropic does not support routing Claude Code to non-Claude models through any gateway.

Sourcecode.claude.com/docs/en/llm-gatewayread 2026-09-29

The Claude Agent SDK gives the same tools, agent loop and context management that power Claude Code as a Python and TypeScript library, with hooks, subagents, MCP, permissions, sessions, skills and plugins, and the CLI can be driven from any language with -p and JSON output.

Sourcecode.claude.com/docs/en/agent-sdk/overviewread 2026-09-29

claude plugin eval runs a suite of realistic prompts in fresh, isolated non interactive Claude Code sessions, grades each run with regex, tool call or model judged rubric graders on the reply, transcript or files, compares against a no plugin baseline, reports pass rates per case, and gates CI on the score, so a team can catch regressions when its setup or the model changes; /verify and /code-review check individual changes.

Sourcecode.claude.com/docs/en/plugin-evalsread 2026-09-29

The Chrome integration lets the agent operate a visible browser session with the user's logins, navigating, clicking, typing, filling forms, uploading files, reading console errors and DOM state and taking screenshots, in Chrome, Edge and other Chromium browsers, with site permissions set in the extension and a pause for logins and CAPTCHAs; computer use extends this to native macOS apps.

Sourcecode.claude.com/docs/en/chromeread 2026-09-29