Give your AI agent safe access to GitHub

Open the AgentBlocks dashboard, click “Connect GitHub,” and authenticate. That's the entire setup. Your agent now has access to your repositories — but only within the rules you're about to set. You can disconnect at any time.

Choose exactly what your agent can do on its own — browsing repos, reading commits, viewing PRs — and what needs your say-so. Require approval before it pushes code or opens PRs. Block anything you don't want it to touch at all. You control every action, on every repo.

Open Claude Code or your terminal-based AI agent and add AgentBlocks as an MCP server. One command. Your agent instantly knows it has GitHub access and exactly what it's allowed to do.

Ask your agent to work on your codebase and it gets started. It browses the repo, reads the right files, makes edits — all the read and write operations that you've set to auto-allow happen without interruption. When it hits something that needs a push or PR, it stops and asks.

When your agent is ready to push code or open a pull request, it shows you exactly what it wants to do — the repo, the branch, the files changed, the commit message — before anything is pushed. One tap to approve, and it's done.

Every approved action uses a token issued for that moment only. It can push to that one branch, in that one repo — nothing else. It expires in minutes. Once the action completes, it's gone. No credentials sitting around, no way for them to be reused or leaked.

Once you're happy with your rules, your agent works through your codebase on its own — reading files, making edits, building up context. It only asks when there's something to push or merge. You approve from wherever you are in seconds, and it keeps going.