Skip to content

agent-mcp

agent-mcp: a lavender and white plug beside the project name, its tagline and four pillars: one manifest, dry-run preview, sops and age secrets, verified in CI

Declare every MCP server once and install it into Claude Code, Codex, OpenCode and Pi on macOS, Arch and Debian.

What it does

  • One manifest. servers.yaml declares every MCP server and the dependencies it needs. servers.schema.json defines the format, and task manifest:check validates it.
  • Installs into four harnesses. scripts/installer.py registers every server into each harness found on PATH: Claude Code, Codex, OpenCode and Pi.
  • Dry-run preview. task install -- --dry-run and task sync -- --dry-run print the plan per harness and write nothing, not even the state file or backups.
  • Undo. task uninstall removes only the entries this repository recorded in its state file. Entries added by hand are never changed.
  • Secrets with sops and age. secrets/secrets.sops.yaml holds the Telegram API id and hash, encrypted. Both tools run in one-off containers, so a host needs only Docker.
  • Verified in CI. The manifest, the installer tests, the secrets store and the Compose file are checked on every push and pull request. See CI and brand.

Where to go next

  • Getting started: prerequisites and the first install.
  • Operator guide: setup per machine, day-to-day commands, undo and troubleshooting.
  • Servers: the servers the manifest declares.
  • CI and brand: what CI checks and where the brand assets come from.