Configure TOMC in Cursor or TRAE

中文 · Assistant setup · Validation record

Use TOMC's local stdio MCP server with either client. Checks so far cover the TOMC server and configuration generation. Cursor and TRAE installation and tool calls have not yet been verified on a real machine.

The quickest route, from a checkout with .[mcp] installed (Python 3.10 or newer), is python -m tomc setup --client cursor --write (or --client trae --write --config <file>), then python -m tomc doctor. --write backs the file up first, keeps your other servers and can be run again safely. The manual route below uses the locked runtime and needs no Python environment of your own.

Use the locked runtime

Install UV and make uv available on PATH. Clone the public TOMC repository:

git clone https://github.com/ZipengWu365/TOMC.git

The plugins/claude-code folder contains the shared TOMC source, pyproject.toml, uv.lock and src/server.py. Its runtime can also be registered as a local MCP server in other clients. No Claude Code installation or extra model API key is needed for this route. First startup can download compatible Python and locked dependencies through UV.

Before connecting, download the locked dependencies with your actual runtime directory:

uv sync --locked --directory "/absolute/path/to/TOMC/plugins/claude-code"

Copy the template for Cursor or TRAE. The Cursor template explicitly sets type to stdio, as required by its current configuration reference. Replace both absolute-path placeholders: the runtime directory and its src/server.py file. Use forward slashes in Windows JSON paths, for example C:/Tools/TOMC/plugins/claude-code. Keep the checkout in place; update both paths after moving it. If UV is outside PATH, replace command with its actual executable path. TRAE's documented command restriction excludes spaces in that executable path; directory/script paths belong in separate args strings.

Cursor

Merge mcpServers.tomc into the project's .cursor/mcp.json or your personal ~/.cursor/mcp.json, preserving existing servers. Restart Cursor and check Customize → MCPs for the tomc server and its twelve tools, then say Try TOMC Memory. Official MCP configuration.

For an existing Python environment installed from TOMC source with .[mcp], python -m tomc setup --client cursor --write merges the entry into ~/.cursor/mcp.json (with "type": "stdio"), and python -m tomc setup --client cursor --link prints an Add to Cursor link instead. Both register that environment; they do not install TOMC or dependencies on another computer.

TRAE

In the international client, open Settings → MCP, choose the local runtime where a local/cloud choice appears, and add the template through manual JSON configuration. A project configuration uses .trae/mcp.json; first enable project MCP in Settings → MCP. Preserve other server entries. Official references: international manual/project configuration and Chinese configuration. Menu labels can vary by client edition/version.

With .[mcp] installed from TOMC source, python -m tomc setup --client trae prints an alternative JSON configuration using that environment's absolute Python path, and python -m tomc setup --client trae --write --config <path to the file> merges it into TRAE's settings file or a project .trae/mcp.json after a backup. If this option is unavailable, update the source checkout or use the UV template above. A Python executable path containing spaces conflicts with TRAE's documented command restriction; use the UV template with uv on PATH instead. --link remains Cursor-only.

Notebooks and feedback

Notebooks default to local plaintext SQLite at ~/.tomc/memory.sqlite3. To choose a different file, add env alongside command and args:

{
  "env": {"TOMC_MEMORY_PATH": "/absolute/path/to/tomc-memory.sqlite3"}
}

Clients sharing that path share explicitly saved notes. TOMC does not collect other chats or provide cloud synchronization. The twelve tools in 0.3.1 are try_tomc, prepare_context, remember_memory, recall_memory, list_memories, forget_memory, erase_all_memory, get_status, get_usage_stats, list_contexts, get_context and make_handoff. Notes are stored only when you ask; prepare_context and recall_memory results are not saved unless you turn that on with tomc settings --save-mcp-results on. Delete with forget_memory, or everything with erase_all_memory / tomc erase-all --yes. Returned memory enters the host assistant's context and follows its data policy. A tool call does not remove history already held by the host.

Please report setup feedback through GitHub issues, including the client/edition, version, OS, configuration route and error message. Omit API keys, private history and notebook contents.