Claude Agent SDK — Anthropic's Framework for Production AI Agents

Claude has no persistent state between conversations. Every API call is stateless. The Claude Agent SDK solves this by giving Claude a proper agent runtime — a loop that keeps running, executes tools, manages context, spawns subagents, and connects to external systems via MCP.

The SDK exposes the same infrastructure that powers Claude Code. Not a simplified version — the actual runtime, made programmable.

History

The Claude Agent SDK started life as the Claude Code SDK, announced May 22, 2025 alongside Claude Opus 4. The original purpose was narrow: let developers run Claude Code headlessly in CI/CD pipelines. As usage expanded beyond coding — email assistants, research agents, monitoring bots — Anthropic renamed it to the Claude Agent SDK on September 29, 2025.

The name change reflects the broadened scope. The tooling is identical.

Installation

TypeScript:

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

Python:

pip install claude-agent-sdk

Set your API key:

export ANTHROPIC_API_KEY=your-api-key

Core Capabilities

Built-in Tools

The SDK ships with a full tool suite — no custom implementation required:

Tool	What it does
Read	Read any file in the working directory
Write	Create new files
Edit	Make precise edits to existing files
Bash	Run terminal commands, scripts, git operations
Glob	Find files by pattern
Grep	Search file contents with regex
WebSearch	Search the web for current information
WebFetch	Fetch and parse web page content

You specify which tools the agent can access. Read-only agent? allowedTools: ["Read", "Glob", "Grep"]. The agent cannot use tools outside this list.

Sessions and State

By default, each query() call starts fresh. Capture the session ID and pass it back to resume where you left off — including full file-reading history, conversation context, and analysis already done.

This enables multi-step workflows where an agent reads a codebase in one session, then modifies it in another without re-reading everything from scratch.

Subagents

Pass Task in allowedTools and define named agent configurations in the agents option. The main agent can then spawn these specialized subagents via the Task tool — they run in isolated context windows and report back to the orchestrator.

This is how you build parallelized pipelines: a research subagent and a writing subagent working simultaneously, with a coordinator merging results.

MCP Integration

Any MCP server — database connectors, browser automation, file systems, APIs — integrates in five lines:

options: {
  mcpServers: {
    playwright: { command: "npx", args: ["@playwright/mcp@latest"] }
  }
}

The agent gets MCP tools added to its available toolset automatically.

Hooks

Hooks fire at key lifecycle points: PreToolUse, PostToolUse, Stop, SessionStart, SessionEnd. Use them to log file modifications to an audit trail, validate tool calls before execution, or send alerts when the agent stops.

Budget Controls

Set max_budget_usd in your agent options. When the running API cost exceeds this threshold, the agent stops. Essential for production deployments where runaway agents could generate unexpected API bills.

Pricing

The SDK itself is MIT-licensed and free. Costs come from Anthropic API usage per token:

• Claude Haiku 4.5 — $1 input / $5 output per million tokens. Good for lightweight agents with simple tasks.
• Claude Sonnet 4.6 — $3 input / $15 output per million tokens. The default choice for most agent work.
• Claude Opus 4.6 — $5 input / $25 output per million tokens. For complex reasoning tasks where quality matters more than cost.

Prompt caching (for repeated context like system prompts or large documents) cuts input costs up to 90% on cache hits. For long-running agents that repeatedly reference the same codebase or documents, caching is significant.

Backend Options

The SDK supports Anthropic’s API directly, but also works through:

• Amazon Bedrock — set CLAUDE_CODE_USE_BEDROCK=1
• Google Vertex AI — set CLAUDE_CODE_USE_VERTEX=1
• Microsoft Azure AI Foundry — set CLAUDE_CODE_USE_FOUNDRY=1

Teams already in AWS or Google Cloud infrastructure can route API calls through their existing cloud accounts rather than direct Anthropic billing.

What It Doesn’t Include

No container isolation. The agent has access to your filesystem unless you sandbox it externally. If you’re running untrusted prompts or want production safety, you need to add Docker or similar yourself. NanoClaw wraps the Agent SDK with OS-level container isolation — if that’s a concern, start there.

No managed hosting. You deploy, you manage. There’s no Anthropic-hosted endpoint that runs your agents for you.

No built-in multi-LLM support. The Agent SDK is Claude-only. For workflows that mix Claude with other models, you’d build that routing layer yourself.

Our Take

The Claude Agent SDK is the serious option for Claude-based agents. It’s not a wrapper or a convenience layer — it’s the actual runtime, open-sourced. The built-in tools, the subagent architecture, the MCP integration, and the budget controls are all production-grade.

The tradeoff: you bring the infrastructure. Sandboxing, hosting, monitoring, deployment — that’s your problem. The SDK gives you the agent engine; everything around it is yours to build.

Best for: Developers building automated pipelines, research agents, coding automation, and multi-agent systems where Claude is the core AI engine.

Skip if: You want a ready-to-run platform — look at NanoClaw (container isolation + messaging integration out of the box) or OpenClaw (simpler self-hosted agent runtime with a large community).

Rating: 9.0/10