mcp-sync
mcp-sync
A project's MCP servers are recorded in one doc — MCP.md at the repo root — and you mirror them
into the three agents' project-local configs. There is no script; you do this by following the steps
below. It's safe by hand because these are small files: read one whole, change only its MCP section,
write it back — except Claude Code, which you drive through its CLI (see below).
Scope is deliberately narrow: project-level only, three targets (Claude Code, opencode, Codex). No global/user configs, no other tools.
Nothing MCP-related is committed. MCP.md, opencode.json, and .codex/config.toml are all
gitignored; Claude Code's servers live in local scope inside ~/.claude.json, which is already
outside the repo. So the whole MCP setup stays private to this machine — the manifest is documentation,
not a shared artifact.