Taichu MCP

Start with the local executable

These examples require a Taichu executable. The package for GWC for VS Code is still being prepared; check the Download page for verified release availability.

taichu --mcp-serve

Let your MCP client launch the process. Set its working directory to the project you want to use: relative file paths and Taichu settings resolve from that directory. Startup diagnostics go to standard error; standard output carries the protocol.

Available tools

• read_file

  Read a local file.

• write_file

  Create or replace a local file.

• edit_file

  Apply a text replacement in a local file.

• notebook_edit

  Edit cells in a notebook.

• glob

  Find files by path pattern.

• grep

  Search file contents.

• prior_art

  Search the codebase for existing implementations.

• report_findings

  Validate structured findings. Standalone MCP cannot persist them; see the limits below.

Connect from Codex

Add this table to ~/.codex/config.toml, replacing both example paths with your actual executable and project directory.

[mcp_servers.taichu]
command = "C:/path/to/taichu.exe"
args = ["--mcp-serve"]
cwd = "C:/path/to/project"

On other platforms, use the path to your Taichu executable. Reload the client after changing its configuration, then inspect its MCP server list. See the official Codex MCP documentation for client-specific options.

Other MCP clients

For clients that accept an mcpServers JSON object, use the same executable and flag. Start the client in your project directory, or set the server working directory using that client's configuration.

{
  "mcpServers": {
    "taichu": {
      "command": "C:/path/to/taichu.exe",
      "args": ["--mcp-serve"]
    }
  }
}

Current limits

• This server does not provide shell execution, browser automation, media generation, agent orchestration, or scheduled jobs.
• report_findings validates the input but returns persistence_unavailable because durable findings require an owned Taichu session.
• Local tools run with the permissions of the launched process. Taichu loads the working directory's configured hooks; a pre-tool hook can block a call.
• The connected AI client remains responsible for its model connection and model usage charges.

Troubleshooting

If the server does not appear, check the executable path, the --mcp-serve flag, and the client's startup error log. If a file cannot be found, check the configured working directory. A blocked tool result may come from a configured Taichu hook.

For cloud API requests, see the API reference. To connect third-party tools to GWC, return to the MCP overview.