Sangam-Git

Connect AI agents

Sangam-Git is built for a world where AI coding tools write a large share of the code. You can connect Claude Code, Cursor, Copilot, Codex, Windsurf, Aider — anything that speaks MCP or webhooks — and Sangam-Git will label what they produce and keep a human in charge.

Why this is different

On other hosts, a commit from an AI agent looks exactly like a commit from a person. Nobody can tell what a bot wrote. Sangam-Git changes that: work from an AI agent is labelled at the source, and AI-authored pull requests cannot merge without a human's approval. This isn't a setting someone can forget — it's a rule in the system.

Create an agent token

An AI agent authenticates with an access token, just like you do — but a special kind.

  1. Go to Settings → Tokens.
  2. Give the token a name (e.g. claude-code).
  3. Choose For an AI agent. This is the important part: every commit and pull request made with this token is automatically tagged "AI agent."
  4. Copy the token now — it's shown only once.

Because the token is marked as an agent, reviewers always know which changes came from a tool, so review effort goes where the risk actually is.

Connect over MCP

Sangam-Git exposes a single MCP endpoint that any MCP-capable agent can use. For example, with Claude Code:

claude mcp add --transport http sangam \
  https://sangam.example.com/mcp \
  --header "Authorization: Bearer sgm_your_agent_token"

One endpoint, one command. The same tools and the same rules apply to every agent — none of them "owns" your Git host.

Let the agent work

Now your agent can read the repo, create branches, commit, and open pull requests — all through Sangam-Git. Everything it pushes is labelled. When it opens a PR, the PR is marked AI-authored and the human-approval gate applies automatically.

Bring your own agent (webhooks)

You may prefer to run your own review agent on a plan you already pay for. Sangam-Git supports that too: it POSTs signed JSON to your webhook URLs on pull-request and issue events. Your CI receives the event, runs your agent (Claude Code, Cursor, …), and posts the review back — at zero model cost from us.

Set this up under a repository's Settings → Webhooks. Always verify the X-Sangam-Signature-256 header with your signing secret so you know the event really came from Sangam-Git. (See Security.)

What Sangam-Git's own AI can do

If you'd rather use the built-in AI features, they run only when you ask, on the specific diff involved, and each run shows the model used and its cost:

  • Summarise a pull request in plain language.
  • AI review — bounded, high-confidence findings only (real bugs, not style nits).
  • Explain config changes — turn a confusing YAML/JSON diff into "adds a discount field to the Orders form."

You can turn all AI off for your whole account or organization. When off, your code is never sent to any model — while labels, risk lanes, and secret scanning keep working locally. (See our Trust & data page.)

The principle: AI can propose as much as it likes. A person still decides what merges into main.