Claude Agent SDK for Python

When Anthropic renamed "Claude Code SDK" to the Claude Agent SDK, it signaled where the value really lives: not in the chat loop, but in turning Claude Code's agent harness — its tool execution, permission gating, and session management — into a library you call from your own code. You don't reimplement an agent; you drive the same one Claude Code runs, with deterministic hooks where it matters.

What Sets It Apart

• Two interaction shapes, one engine. query() is a stateless async generator for fire-and-forget tasks; ClaudeSDKClient keeps a bidirectional session for interactive, multi-turn work with mid-stream control.
• In-process MCP tools. Decorate a Python function with @tool and it becomes a callable tool with no separate process or transport — closing the latency and ops gap that external MCP servers impose.
• Hooks make behavior deterministic. PreToolUse/PostToolUse callbacks let you inspect, rewrite, or block actions (e.g., veto a Bash command) instead of relying on the model to behave.
• Real permission surface. allowed_tools allowlists, permission_mode, and a can_use_tool callback give code-level control over what the agent can touch.

Who It's For

Great fit if you want production agents that reuse Claude Code's battle-tested file/Bash/edit tools and need guardrails enforced in code rather than prompts. Look elsewhere if you only need plain text completions — the SDK bundles and manages the Claude Code CLI under the hood, which is overhead you don't need for a simple chat call. Note it targets Python 3.10+ and the agent runs through that bundled CLI.