MCPClerk

Manage MCP configuration for Claude Code and Codex. MCPClerk is a local CLI for inspecting configured servers, editing native host settings, keeping an operation catalog, and checking task-scoped MCP calls.

MCPClerk edits configuration; Claude Code and Codex still run MCP servers and own runtime approval and trust. It does not proxy tool traffic or launch servers during inventory. A listed entry is configured state, not proof of a live connection.

Install

MCPClerk requires Python 3.10 or newer. From a checkout:

python -m pip install .
mcpclerk --help

For development, install .[dev] and run python -m pytest.

Related MCP server: mcp-policy-gateway

Start here

list inspects native configuration without launching MCP processes. Its inventory omits commands, URLs, environment values, and headers:

mcpclerk list
mcpclerk list --host codex --project /path/to/project

To add a server, put its native definition in a JSON file and choose a host and scope. For example, server.json can contain a stdio definition:

{"command":"npx","args":["-y","@modelcontextprotocol/server-filesystem","/path/to/project"]}

Then add it to a user configuration:

mcpclerk add files --host claude --scope user --file server.json

The command reports an undo transaction ID. Use mcpclerk undo ID to restore the previous file if it has not changed since. See the command guide for project scopes, per-tool settings, catalog policy, and task judging.

Catalog policy is separate from host approval. mcpclerk judge reads one JSON request from standard input and returns a classification and decision without echoing its arguments. Keel calls it for MCP tool events; the decision uses declared task permissions and project target policy.

Claude Code and Codex remain the runtime authorities for native configuration and project trust. Managed settings, project trust, session flags, and live connections can change effective availability beyond what a local file inventory can establish.

License

MIT.