anthropics/claude-agent-sdk-python
Claude Agent Sdk Python
Claude Agent SDK is a Python library from Anthropic for building AI agents that use Claude’s tool‑calling, streaming, and session‑management capabilities. It provides a structured framework for intercepting and logging every tool call an agent makes, managing conversation sessions across multiple turns, and controlling which tools the agent is allowed to invoke through configurable permission rules.
The SDK distinguishes itself with middleware‑driven tool interception that lets developers block, modify, or require approval for tool calls before or after execution. A permission policy engine evaluates each tool request against allowlists, deny patterns, and baseline permission modes—automatically approving, denying, or delegating decisions. Session lifecycle hooks fire custom code at start, stop, or specific conversation events, giving fine‑grained control over state and cleanup.
Beyond governance, the SDK supports real‑time streaming of agent responses, both for prompts and tool results, enabling incremental processing before a full reply completes. Developers can define custom tools with typed JSON schemas that the agent can discover and invoke, and run interactive sessions with custom extensions that include lifecycle hooks. An in‑process MCP server and a lightweight local server allow external clients to discover and call tools from the same process.
Features
- Claude Integration Frameworks - Delivers a comprehensive Python SDK for building Claude-based agents with tool governance, streaming, and lifecycle management.
- Typed JSON Schemas - Provides typed JSON schema definitions for custom tools that agents automatically discover and invoke.
- Tool Access Controls - Controls agent tool access through allowlists and permission modes for automatic approval or denial.
- Agent Response Streamers - Streams agent responses incrementally for real-time processing.
- Agent Governance - Intercepts, logs, and modifies tool calls to enforce governance policies.
- Cross-Session Conversation Memories - Persists context across exchanges to continue conversations with memory.
- In-Process MCP Server Implementations - Runs an in-process MCP server to expose tools to external clients.
- Agent Response Streams - Ships streaming agent response capabilities for real-time display and progressive processing of prompts and results.
- Tool Call Interception Middleware - Passes tool calls through a middleware chain that can block, modify, or approve execution.
- Agent Workflow Interception - Middleware hooks intercept and modify AI agent tool calls before or after execution.
- Streaming Response Architectures - Streams agent response tokens incrementally before full reply completes.
- Tool Permission Engines - Evaluates tool access against allowlists, deny patterns, and permission modes.
- Allowlist and Deny Rule Middleware - Implements middleware-based tool permission control with allowlists, deny patterns, and approval delegation.
- Agent Session Lifecycle Hooks - Ships session lifecycle hooks that fire custom code at agent session start, stop, and conversation events.
- Agent Tool Interception - Ships a library for intercepting and gating AI agent tool calls with pluggable middleware chains.
- Lifecycle Hooks - Attaches lifecycle hooks to register callbacks at specific conversation phases.
- MCP Tool Call Loggers - Logs all tool call activity via MCP for auditing, debugging, and compliance.
- Custom Extension Integrations - Runs interactive sessions with custom extensions in-process, supporting lifecycle hooks.
- Tool Allow and Deny Patterns - Defines pattern-based allow and deny rules that automatically approve or block specific agent tool requests.
- Custom Tool Definitions - Defines custom tools as functions with typed schemas and descriptions for callable use.
- Tool Schema Definitions - Defines custom tools using JSON Schema for typed parameters and descriptions.
- One-Shot Query Execution - Sends one-off queries to new sessions without maintaining state.
- Real-Time Text Streaming - Streams messages in real time for incremental input and output.
- Incremental Response Clients - Provides a client that streams incremental agent responses and tool results for real-time consumption.
- Agent Permission Modes - Sets permission modes that select baseline approval policies for all agent tool requests.
- Infrastructure and SDKs - Official SDK for building agents with path management and integration.
