MCP

Two MCP servers

Local: agit mcp (stdio)

Runs on your machine, registered per-runtime by agit setup. The agent you’re talking to uses it without leaving the runtime:

Tool	What it does
search	search the readable corpus — “has someone done this before”
show	read a session (ref, ref#n, ref#n.k)
view	a branch’s VIEW composition
status	who am I + sync state
commit	settle the current session (--milestone supported)

Each tool is the CLI command, wrapped — there is no second implementation to drift. Search is intentionally MCP-first: an agent that’s stuck should query the corpus inline rather than you alt-tabbing to a terminal.

Remote: https://<hub>/mcp (streamable HTTP)

The hub itself is an MCP server. Point any MCP client at it — no agit install required — to read the public corpus, or your private agents too with a token in an Authorization: Bearer <token> header (the token agit login stored in ~/.agit/credentials.json):

https://agent-git.com/mcp

Same read tools, server-side: search across the hub, fetch an agent’s sessions, render a turn. This is how “give my agent access to the team’s agent history” works without giving everyone a laptop setup.

Skills

agit setup drops the agit usage note where each runtime keeps its instructions (see Supported runtimes for the paths). The note teaches three things:

Self-reference. @ means your own line. agit log, agit show @#3, agit commit @ all work from inside a session because AGIT_SESSION is set.

Milestone discipline. When a phase truly completes, settle it:

agit commit --milestone "short phase summary" --tag ms-name --code

The merge-agent protocol. When an agent is resumed with AGIT_MERGE_TX set, it is the merge agent: read both sides (agit view <source>, agit show <source>#n.k), pick turns, reconcile shared files (cd $(agit repo path)), write the summary, then agit merge --continue. Recon moves stay out of the VIEW — explore freely, don’t pollute the context that follows.

And one prohibition: no rebase, no amend, no force-push. History only grows.