Claude Agent SDK
The Claude Agent SDK is a software development kit provided by Anthropic to help developers build custom AI agents based on the Claude Code Agent Harness framework. The SDK supports both Python and TypeScript, providing standardized interfaces and toolchains that enable developers to extend and customize AI agent behaviors and capabilities. It is applicable to scenarios such as automated task execution and workflow orchestration.
Core Facts
Agent Skills是Anthropic发布的能力封装标准，定义为一个包含Skill.md文件的目录，可包含说明、脚本和其他资源
The core interface of the Agent SDK is a query() function; passing in a prompt and configuration starts the agent loop, and key configuration items include settingSources, allowedTools and systemPrompt.
Claude Agent SDK supports multi-step reasoning, tool calls, and environment interaction, including capabilities such as code writing, file operations, and command execution.
Timeline (last 90 days)
The Claude Agent SDK package bundles the native Claude Code binary, so most machines need no additional installation.
The claude agent sdk tutorial recommends using the uv package manager, and the Python version must be 3.10 or above.
In the Claude Agent SDK, deny rules can intercept the rm command and take effect in all modes (including bypassPermissions), but deny rules match the literal command, and forms like /bin/rm can bypass them, so a pre-tool-use hook is needed to check every bash call at the front of the chain.
In the Claude Agent SDK, an uncaught tool exception does not terminate the agent loop; the server converts it into an error result carrying the original exception information.
The Claude Agent SDK by default uses tool search to lazily load MCP tool schemas, to avoid a large server occupying the context window from the start.
In the Claude Agent SDK, the fully qualified name of an MCP tool is mcp__servername__toolname (joined with double underscores), and this full string must appear in the allowed list exactly as is, otherwise the call falls to permission mode.
Common values of the Claude Agent SDK permission mode include default (calls not pre-approved require callback confirmation, and no callback means denial), acceptEdits (automatically accept file edits), dontAsk (deny if not pre-approved), and bypassPermissions (skip all permission checks).
In the Claude Agent SDK, the check order for a tool request is: hooks → deny rules → ask rules → permission mode → allow rules → can use tool callback (in default mode, no callback means denial).
In the Claude Agent SDK permission system, the allowed list only means pre-approval; tools not listed are not turned off but instead go to permission-mode handling; to remove a tool you must put its bare name in the disallowed list.
The Claude Agent SDK does not automatically load .env files; you need to set the ANTHROPIC_API_KEY environment variable in the shell running the agent, or manually load it via dotenv or similar before the SDK starts.
40 more timeline events
All Facts (20)
Agent Skills是Anthropic发布的能力封装标准，定义为一个包含Skill.md文件的目录，可包含说明、脚本和其他资源
80%VerifiedThe core interface of the Agent SDK is a query() function; passing in a prompt and configuration starts the agent loop, and key configuration items include settingSources, allowedTools and systemPrompt.
65%VerifiedClaude Agent SDK supports multi-step reasoning, tool calls, and environment interaction, including capabilities such as code writing, file operations, and command execution.
65%VerifiedThe Claude Agent SDK has built-in tools such as Web Search and Web Fetch, giving the Agent the ability to search the internet and scrape web pages.
65%Unverifiedclaude-memory-compiler采用三层自动化流水线架构：会话捕获层（Hooks机制）、知识提取层（Claude Agent SDK）和知识编译层（LLM Compiler）
90%Unverified该方案的所有数据和推理都在本地完成，不会外传到第三方服务器，适合对数据隐私有严格要求的企业场景
90%Unverified通过设置环境变量ANTHROPIC_BASE_URL指向LiteLLM Proxy的本地端口，可以将Claude Agent SDK的请求重定向到本地模型，无需修改SDK源码
90%Unverified2024-2025年，Anthropic Claude Agent SDK、OpenAI Agents SDK、Google ADK等Agent框架集中涌现
85%UnverifiedClaude Agent SDK提供工具注册与调用、多轮对话上下文管理、智能体间消息传递与协调机制以及结构化输出控制等能力
85%Unverified框架能力与推理能力解耦是AI工程领域的重要趋势，Claude Agent SDK提供Agent编排框架，实际推理可由任何兼容模型完成
75%Unverified该方案的整体架构分为三层：Agent Framework层（任务管理和调度）、Claude Agent SDK层（核心Worker）、LiteLLM Proxy层（请求代理转发）
60%UnverifiedThe solution of connecting the Claude Agent SDK to a local large model via LiteLLM Proxy can reduce API call costs from nearly $1 per simple task to almost zero.
60%UnverifiedThe Claude Agent SDK package bundles the native Claude Code binary, so most machines need no additional installation.
50%UnverifiedThe Claude Agent SDK does not automatically load .env files; you need to set the ANTHROPIC_API_KEY environment variable in the shell running the agent, or manually load it via dotenv or similar before the SDK starts.
50%UnverifiedThe Claude Agent SDK by default uses tool search to lazily load MCP tool schemas, to avoid a large server occupying the context window from the start.
50%UnverifiedIn the Claude Agent SDK, the check order for a tool request is: hooks → deny rules → ask rules → permission mode → allow rules → can use tool callback (in default mode, no callback means denial).
50%UnverifiedIn the Claude Agent SDK permission system, the allowed list only means pre-approval; tools not listed are not turned off but instead go to permission-mode handling; to remove a tool you must put its bare name in the disallowed list.
50%UnverifiedCommon values of the Claude Agent SDK permission mode include default (calls not pre-approved require callback confirmation, and no callback means denial), acceptEdits (automatically accept file edits), dontAsk (deny if not pre-approved), and bypassPermissions (skip all permission checks).
50%UnverifiedIn the Claude Agent SDK, the fully qualified name of an MCP tool is mcp__servername__toolname (joined with double underscores), and this full string must appear in the allowed list exactly as is, otherwise the call falls to permission mode.
50%UnverifiedIn the Claude Agent SDK, an uncaught tool exception does not terminate the agent loop; the server converts it into an error result carrying the original exception information.
50%