Use with your agent

You can ask your agent to use AgentGit in plain language. Skills teach the workflow; MCP gives the agent tools to search, inspect, and save history. Complete Get Started first if you want to record your own conversations.

Set up your agent

agit setup

This installs the AgentGit skill, available runtime hooks, MCP registration, and the project’s AGENTS.md integration. To target one runtime, use agit setup --runtime codex, for example. Reload your runtime if it needs a restart to discover new tools or skills.

Then try:

Use AgentGit to save this conversation with the milestone “API design agreed”.

Search our AgentGit history for this error, read the relevant turns, and explain which solution actually worked.

Use AgentGit to compare our current approach with the alternative session. Propose which decisions we should keep before merging.

See Supported runtimes for installation paths and hook availability. Saving locally and publishing to the Hub are separate actions; ask for a push explicitly when you want to publish.

Local MCP server

agit mcp runs on your machine over stdio and is registered by agit setup. Your agent uses these tools from within its runtime:

Tool	What it does
search	Search hosted history you can read; requires sign-in
show	Read a conversation, turn, or event by reference
view	Inspect a branch’s VIEW composition
status	Inspect identity and sync state
commit	Save the session supplied by AGIT_SESSION, optionally with a milestone
rc_status	Inspect the local remote-control daemon’s connection and supervised sessions
rc_list	List Cloud devices available to the account

The tools call the CLI implementation. If an operation needs an option the MCP tool does not expose, the agent can use the corresponding CLI command with the installed skill’s guidance.

The local commit tool inherits AGIT_SESSION from the MCP server process; it does not accept a target argument. If that environment value is missing or stale, ask the agent to use agit commit owner/repo@branch in its shell with the intended target. Importing a conversation does not update the environment of an already-running MCP server.

Remote MCP server

The Hub provides a read-only MCP endpoint for clients that do not need a local agit installation:

https://agent-git.com/mcp

Configure this as a Streamable HTTP MCP server in your client. Public repository browsing works without authentication. search, session_hits, and private repository access require a valid token from the same Hub in an Authorization: Bearer <token> header. Private content also requires a repository read grant. Store the token in your client’s credential configuration.

Tool	What it does
search	Search readable sessions, repositories, PRs, or people
session_hits	Locate matches inside a session using the repository, session ID, and ref from a search result
list_agents	List repositories visible to the caller
get_agent	Fetch repository metadata by owner and name
list_sessions	List sessions at a ref; ref="*" lists across the repository; retain returned cursors for later pages
read_turns	Read a bounded conversation window; from is inclusive and to is exclusive

Each POST is an independent JSON-RPC request. The endpoint supports initialize, ping, tools/list, and tools/call, with no persistent MCP session or SSE stream. The tools use the Hub’s existing access controls and result-size limits. Remote MCP does not expose the local server’s commit tool.

Search filters

The local and remote search tools accept author, since, and before as separate arguments for session searches. These are not query qualifiers:

{
  "author": "alice@example.org",
  "since": "2026-09-01",
  "before": "2026-10-01",
  "page": 1,
  "limit": 20
}

Omit query to search using only these filters. author matches the selected saved version’s Git author name or email exactly, ignoring case. It is recorded Git metadata, not an authenticated account or repository owner. since includes the boundary and before excludes it. Time refers to the saved version’s Git committer timestamp, independently of transcript event timestamps. Date-only boundaries mean midnight UTC; RFC3339 timestamps with offsets normalize to UTC.

applied_filters echoes the normalized predicates. The local tool rejects a Hub response that omits or changes the requested filters. incomplete: true means that hits and total are lower bounds, including when metadata could not be read; filters are never silently relaxed. Inspect unknown for unsupported query qualifiers. Use page and limit to continue through results. The local tool also accepts queries for a batch with shared filters; the remote tool accepts one search per call.

What the skill teaches

The installed skill includes command references that the agent reads when needed. It teaches the agent to:

• Run agit status and select the correct repository before creating or importing a session.
• Use @ only when AGIT_SESSION selects the intended session; otherwise use an explicit owner/repo@branch target.
• Save completed work with a milestone, and add --code only when a code commit is also intended.
• Reconcile forks by reading their VIEWs, selecting useful turns, reviewing shared-file changes, and writing a merge summary before continuing.
• Preserve conversation history instead of rebasing, amending, or force-pushing it.

The merge-agent protocol

When AgentGit launches a merge agent, AGIT_MERGE_TX identifies the target transaction. Automatic settlement is suspended for that merge session, so its exploratory conversation is not appended to the target’s history.

The agent reads both VIEWs and their events with agit view and agit show, selects context with merge pick and merge drop, and reconciles shared files under the path returned by agit repo path owner/repo@branch. It then writes the required summary and completes the transaction with agit merge --continue. The manual merge workflow shows the commands and their order.

You can use the skill without MCP: the agent can run agit directly. Everyday commands shows the terminal equivalents; Forks and merges covers reconciliation.