Install

Five commands, zero to published

# 1. Install (needs node ≥ 20; the installer fetches the binary and
#    runs `agit setup` for you — skills, hooks, MCP, all of it)
npx -y create-agit

# 2. Sign in. No password in the terminal: a browser approval opens by default;
#    on a browserless box `agit login --device` prints a code to enter on any
#    other device. (CI: read a token from stdin — any access token minted by an
#    interactive `agit login` works, ~/.agit/credentials.json)
agit login                        # browser approval
agit login --device               # prints a code, enter it on the website
agit login --with-token <<< "$AGIT_TOKEN"

# 3. Wire the integrations once (idempotent — safe to re-run)
agit setup

# 4. Adopt a session that's already running. Import settles the turns
#    it has so far and places the session on branch `exp`.
agit import <session-id> -n photo -b exp

# 5. Publish. From now on every turn settles automatically.
agit push photo

After step 3, every user turn ends in a turn commit — you don’t get to forget to save, just like git with code.

The installer covers Linux (x64 and ARM64) and macOS (Intel and Apple silicon). Upgrades arrive on their own: the hub announces new releases, and agit upgrade swaps the binary in place.

Which hub you’re talking to

Out of the box the CLI talks to the public hub, https://agent-git.com. Nothing to configure.

Your password never leaves the machine in plaintext — the hub stores a hash; ~/.agit/credentials.json holds the resulting session tokens locally.

Verify the wiring

agit doctor        # local integrity + per-runtime capability table
agit doctor --check-backend
agit status        # who am I, which session, sync state

Next: Concepts — the four mappings everything else follows.