ChatDev/docs/user_guide/en/modules/tooling/README.md

Tooling Module Overview

DevAll currently exposes two tool binding modes for agent nodes:

1. Function Tooling – call in-repo Python functions from functions/function_calling/, with JSON Schema auto-generated from type hints.
2. MCP Tooling – connect to external services that implement the Model Context Protocol, including FastMCP, Claude Desktop, or any MCP-compatible tool stack.

All tooling configs hang off AgentConfig.tooling:

nodes:
  - id: solve
    type: agent
    config:
      provider: openai
      model: gpt-4o-mini
      prompt_template: solver
      tooling:
        type: function
        config:
          tools:
            - name: describe_available_files
            - name: load_file
          auto_load: true
          timeout: 20

1. Lifecycle

1. Parse – ToolingConfig selects FunctionToolConfig, McpRemoteConfig, or McpLocalConfig based on type. Field definitions live in entity/configs/tooling.py.
2. Runtime – When the LLM chooses a tool, the executor injects _context (attachment store, workspace paths, etc.) for Function tools or forwards the request through MCP.
3. Completion – Tool outputs are appended to the agent message stream and, when relevant, registered as attachments (e.g., load_file).

2. Documentation Map

• function.md – Function Tooling config, context injection, best practices.
• function_catalog.md – Built-in function list with usage notes.
• mcp.md – MCP Tooling config, auto-launch, FastMCP example, security guidance.

3. Quick Comparison

Dimension	Function	MCP
Deployment	In-process Python functions shipped with the backend.	Remote: call an HTTP MCP endpoint. Local: launch a process and talk over stdio.
Schemas	Derived from annotations + ParamMeta.	Provided by the MCP server's JSON Schema.
Context	_context provides attachments + workspace helpers automatically.	Depends on the MCP server implementation.
Typical use	File I/O, local scripts, internal APIs.	Third-party tool suites, browsers, database agents.

4. Security Notes

• Function Tooling runs inside the backend process, so keep functions least-privileged and avoid executing arbitrary shell commands without validation.
• MCP Tooling now has explicit remote (HTTP) and local (stdio) modes. Remote only needs an existing server URL; Local launches your binary, so constrain the command/env vars and rely on wait_for_log + timeouts to detect readiness.
• Tools that mutate attachments or code_workspace/ should respect the lifecycle described in the Attachment guide (Chinese for now) to avoid leaking artifacts.