A coding agent is already good at chaining commands with pipes and filters; what it lacks is knowing that your internal search script exists at all. This publishes that list and gets out of the way, which keeps the tool surface small compared with wrapping every command as its own MCP tool.
A server that publishes your own command-line tools as MCP tool definitions. It never runs anything itself — the definitions only tell the model what is available and how to invoke it, and the agent's existing shell access does the work.
- A JSON catalog file where each entry carries a `name`, a `description` and an optional `usage` hint
- `command` to set the exact string to run when it differs from the tool name
- Comments allowed in the catalog file, so you can annotate it
- `--config <path>` or `MCP_CLI_CATALOG_FILE` to point at a catalog other than the default
- Inspection with `npx @modelcontextprotocol/inspector npx mcp-cli-catalog`
Run it as `npx mcp-cli-catalog` (package `mcp-cli-catalog`, version 0.1.0). No account and no key. Two things it needs from you: a catalog file — `~/.mcp-cli-catalog.json` by default — and an agent that already has shell access, since the server does not execute commands.
One command — npx mcp-cli-catalog
