A GitHub personal access token is shaped for a human, not a task. Hand it to an agent and that agent can reach every repo the token's scopes cover. Rithub mints a token scoped to exactly one repo and one action -- read, pr, or full -- and backs it with a native MCP endpoint, so an agent only ever has as much reach as the one job you gave it.
| Question | GitHub | Rithub |
|---|---|---|
| Token scope | Account-wide (classic PAT) or per-repo but multi-permission (fine-grained PAT) | One repo, one of 3 scopes (read / pr / full), minted per agent task |
Can a pr-scope token merge or change settings? | n/a -- GitHub has no equivalent scope | No -- rejected server-side, verified live |
| Native MCP endpoint per repo | None built in -- third-party bridges only | rithub.dev/mcp, built in |
| Built-in agent that opens PRs against labelled issues | Requires a separate Action + bot account you wire up yourself | Built in -- label an issue "agent", it opens the PR |
| Can an agent escalate its own token's scope? | n/a | No -- token-management tools refuse any session authenticated with an agent token, tested live |
read, pr, or full).Want the agent to work without you babysitting it at all? Label an issue agent on a repo with the agent feature turned on, and it opens a pull request that closes the issue, through the same scope fence -- see rithub.dev/mcp-docs for the full walkthrough.
This doesn't make Rithub's agent tooling bigger than GitHub's ecosystem of agent-building third-party products -- it's a smaller, newer surface. What's real today: a working scope fence enforced server-side (not just documented), a native MCP endpoint, and a built-in issue-to-PR loop, all live-verified end to end against production, not a demo.
Mint a scoped agent token Import a repo from GitHub