Claude Code API Documentation: API Reference, Agent SDK, CLI, and MCP

Find the right official Claude Code API documentation: Anthropic Messages API, Claude Code CLI, Agent SDK, MCP servers, headless mode, authentication, and automation examples.

Claude Code API Documentation

This page is a routing guide for developers searching “Claude Code API reference” or “Claude Code API documentation.” The important distinction is that Claude Code is primarily a CLI and agentic coding tool, while Claude API is the platform surface for model calls from your own application.

30-Second Answer

There is no single “Claude Code REST API” that replaces the official Claude API docs. Choose the documentation surface by what you are building:

What you want to do	Use this official surface	Start here
Call Claude from your app	Anthropic Messages API and client SDKs	Anthropic API docs
Run Claude Code from a terminal or script	Claude Code CLI	Claude Code CLI reference
Build an agent that reads files, runs commands, and edits code	Claude Agent SDK	Agent SDK overview
Connect Claude Code to GitHub, docs, browsers, databases, or internal tools	MCP servers	Claude Code MCP docs
Run one-off non-interactive prompts	`claude -p` / Headless Mode	Headless Mode guide
Deploy managed autonomous agents	Managed Agents	Claude Platform docs

If you are asking “where is the Claude Code API documentation,” the most likely answer is: CLI reference for terminal automation, Agent SDK for programmable agents, Anthropic API docs for model calls, and MCP docs for tool integrations.

Official References

Use these as source of truth:

Reference	What it covers	When to use it
Anthropic API documentation	Messages API, models, streaming, tool use, structured outputs, prompt caching, rate limits, billing, SDKs.	You are building your own app with Claude model calls.
Anthropic API reference	Exact HTTP endpoints, request fields, response fields, errors.	You need endpoint-level details.
Claude Code docs	CLI setup, workflows, permissions, MCP, hooks, commands, subagents.	You use Claude Code as a coding assistant.
Claude Code CLI reference	`claude`, `claude -p`, auth, update, flags, output formats.	You need terminal commands or automation flags.
Agent SDK overview	Python and TypeScript SDKs for Claude Code-style agents.	You want programmable agents with tool execution.
Claude Code MCP docs	Connecting external tools and data sources through MCP.	You need GitHub, docs, browser, database, or internal tool access.

Which API Surface Should You Use?

The word “API” is overloaded. Use this decision table before copying code:

Requirement	Best fit	Why
Generate text in a web app	Messages API	You manage prompts, state, tools, and UX.
Add Claude to a backend service	Anthropic client SDK	Official SDKs handle request syntax and auth.
Ask Claude Code to inspect a repo once	`claude -p`	Lightweight one-off automation.
Build an internal code review bot	Agent SDK	It can run a Claude Code-style agent loop in your process.
Connect Claude Code to GitHub issues	GitHub CLI or GitHub MCP	Use CLI for simple audited commands; MCP for repeated structured workflows.
Let Claude Code read internal docs	MCP or pasted context	Use MCP only when repeated access is worth the permission surface.
Deploy a hosted autonomous agent	Managed Agents	Anthropic-managed sessions and infrastructure.

If you only need a model response, do not use Claude Code as an API wrapper. Use the Anthropic Messages API. If you need file reading, command execution, code edits, permissions, and session behavior, look at Claude Code CLI or Agent SDK.

Authentication and API Keys

For Anthropic API workflows, start with an API key from Console:

export ANTHROPIC_API_KEY="your-api-key"

Then use an official client SDK or HTTP request. The Claude Platform docs show examples in Python, TypeScript, Go, Java, Ruby, PHP, C#, cURL, and CLI.

For Claude Code CLI, authentication can be through supported Claude Code sign-in flows or Console/API usage billing depending on how you start the CLI. For application code and production agents, prefer API key authentication through the official platform docs rather than relying on a personal interactive login.

Security baseline:

Do	Avoid
Store keys in environment variables or a secrets manager.	Commit API keys to Git.
Use separate keys for dev, staging, and production.	Reuse a personal local token in CI.
Set spend limits and monitor usage.	Run headless loops without a budget guard.
Rotate keys after exposure.	Paste secrets into prompts, logs, or docs.

Claude API vs Claude Code CLI

The Anthropic API and Claude Code CLI solve different problems:

Surface	You control	Claude controls
Anthropic Messages API	Conversation state, tool loop, UI, persistence, safety checks, deployment.	Model response generation.
Claude Code CLI	Prompt, repo, permissions, tool access, approval flow.	Agentic coding loop inside the CLI.
Agent SDK	Your app process, auth, hosting, approval logic, custom tools, observability.	Claude Code-style tool use and agent loop.

Use the Messages API when the product is your own app. Use Claude Code when the task is development work in a repository. Use Agent SDK when you want Claude Code-style behavior inside a programmable workflow.

CLI Automation API: `claude -p`

For non-interactive CLI automation, use print mode:

claude -p "Summarize the last 20 commits and list release risks"

Process piped input:

git diff --stat | claude -p "Write a release note summary"

Return structured output:

claude -p "Review this diff and return the top risks" \
  --output-format json

Use permissions for safer automation:

claude -p "Review this repository without editing files" \
  --permission-mode plan \
  --tools "Read,Grep,Glob,Bash" \
  --allowedTools "Bash(git diff *)" "Bash(git status *)" \
  --max-turns 3 \
  --output-format json

For deeper examples, see the Claude Code Headless Mode guide.

Agent SDK API

Use the Claude Agent SDK when you need a programmable agent rather than a one-off CLI command. Official docs describe the Agent SDK as giving Python and TypeScript programs the same tools, agent loop, and context management that power Claude Code.

Install:

npm install @anthropic-ai/claude-agent-sdk

or:

pip install claude-agent-sdk

The Agent SDK is a better fit when you need:

Need	Why Agent SDK fits
Custom app integration	You can call the agent from your own service.
Programmatic sessions	Your code owns the workflow and persistence.
Custom approval logic	You can intercept and control tool use.
Observability	The SDK has docs for cost, usage, hooks, and OpenTelemetry.
Production automation	You can run the agent in your infrastructure with explicit auth and limits.

Important distinction: Anthropic’s client SDK gives you direct model API access; you implement the tool loop yourself. The Agent SDK gives you Claude Code-style tool execution and agent behavior.

MCP API and Tool Integrations

MCP is not the same as the Anthropic Messages API. MCP is the protocol Claude Code uses to connect to tools, APIs, databases, docs, and external systems.

Use MCP when:

Claude Code repeatedly needs GitHub issue, PR, or release context;
Claude needs current library documentation before editing code;
a browser, database, monitoring system, or internal API is part of the workflow;
copy-pasting context from another tool is becoming the bottleneck.

Start narrow:

claude mcp add --transport http docs https://api.example.com/mcp

Review:

/mcp

For many GitHub tasks, GitHub CLI integration is simpler and easier to audit than adding a full MCP server. Move to MCP when the workflow becomes repetitive enough to justify the extra permission surface.

Common API Tasks

Code Generation

If you want Claude to generate code inside your own product, use the Anthropic Messages API or client SDKs. You send the prompt, choose the model, manage project context, and decide where generated code appears.

If you want Claude Code to generate code inside a local repository, use the CLI or Agent SDK. Add project rules with CLAUDE.md, then use Plan Mode before approving edits.

Code Explanation

For app-level explanations, send the relevant snippet to the Messages API. Keep the snippet small and avoid sending secrets.

For repository-level explanation, Claude Code is usually better because it can inspect nearby files and project structure:

claude -p "Explain how authentication flows through this repository. Do not edit files." \
  --permission-mode plan

Code Review

For a simple PR summary, use claude -p with git diff. For a production review bot, use the Agent SDK with explicit tools, permissions, cost limits, and observability.

Do not let an automated review agent push changes until its read-only review output is reliable.

Code Refactoring

For small local refactors, use interactive Claude Code with Plan Mode. For repeatable migration workflows across many repositories, consider Agent SDK only after you have a proven manual workflow.

Refactoring is rarely a pure API problem. It usually needs repository context, tests, rollback strategy, and human approval.

API Errors, Limits, and Cost

When API or automation fails, identify which surface is failing:

Symptom	Likely surface	Start here
`ANTHROPIC_API_KEY` rejected	Anthropic API auth	Official API auth docs and Console key settings.
`claude -p` works locally but not CI	CLI auth, PATH, CI env, permission mode	Headless Mode troubleshooting
MCP server does not appear	MCP config, transport, auth, scope	MCP not connecting
Usage or spend rises suddenly	Model choice, loops, long context, MCP output, CI repetition	Claude Code pricing and limits
Endpoint field is rejected	Anthropic API version or request schema	Official API reference.

Before putting a script into CI, add:

explicit auth source;
narrow tool permissions;
--max-turns or budget controls where supported;
logging that avoids secrets;
a failure mode that does not deploy, delete, or push.

Recommended Learning Path

Use the installation guide to make the CLI work locally.
Read Basic Usage for interactive sessions, claude -p, and common commands.
Add CLAUDE.md so Claude understands project rules.
Learn Tools allowlist before granting automation access.
Use Headless Mode for one-off scripts.
Move to Agent SDK only when you are building a real program around Claude Code.
Add MCP servers only when a repeated workflow needs external tool access.

Claude Code API Documentation: API Reference, Agent SDK, CLI, and MCP