Stewards review your PRs from day one — that runs on the GitHub App, so it works no matter which coding agent you use. The in-editor commands (define a steward, clear the backlog) are richest in Claude Code today.
Plugin, MCP auth, and the steward skills. Paste this; it does the rest.
Set up steward.foo for this repository.
Run `/plugin marketplace add contextgraph/claude-code-plugin`, then `/plugin install steward`, then `/mcp` and sign in to mcp.steward.foo when prompted.
Then run `/steward:define-steward` to define my first steward.Connect now and get full PR review today. In-editor commands are on the way.
{
"mcpServers": {
"steward": {
"url": "https://mcp.steward.foo"
}
}
}Complete the browser sign-in, install the GitHub App, and your stewards start reviewing PRs. To define a steward before in-editor commands land, ask your agent to use the steward MCP tools directly.
Setup takes a couple of minutes. Here's what you need on hand — most teams already have all three.
Stewards review pull requests, so point them at a repo where agent-written or human-written PRs actually land. A busy repo shipping dozens of PRs a day is where coverage pays off fastest.
Installing the GitHub App needs org-owner or repo-admin rights. If you don't have them, the browser flow will ask an admin to approve — you can scope the App to a single repo to start and widen it later.
PR review runs entirely on the GitHub App and needs no editor. The in-editor commands — defining a steward, clearing its backlog — are richest in Claude Code today, with Codex, Cursor, and VS Code supported through MCP.
A steward is a long-running owner, not a one-off linter. Once it's connected, coverage compounds with every PR your agents open.
Once the App is installed, every new pull request is reviewed through each steward's mission lens — analytics, observability, security, onboarding, retention, testing, and whatever else you've defined. No configuration files to commit, no CI wiring.
A steward doesn't just leave comments. When it spots a gap in its dimension, it opens a follow-up pull request with the actual fix, so the correction is code you can read, run, and merge — not a thread you have to translate.
Everything a steward wants to change lands in a running backlog scoped to its one dimension of quality. You move from in the loop (reviewing every diff) to on the loop (setting the taste and adjudicating what ships).
No. PR review is fully driven by the GitHub App and works with any agent — or none. Claude Code and other MCP clients add the in-editor commands for defining stewards and clearing backlogs.
Start with one narrow owner for the dimension you worry about most, then add more. Each steward owns a single quality dimension across the whole codebase, so a fleet of them gives coverage no one general reviewer can match.
Yes. Install the GitHub App on a single repo to evaluate, then expand the installation to more repos in your org whenever you're ready.
Reach us at ben@contextgraph.dev or in the community Discord linked in the footer. The docs and MCP endpoint at mcp.steward.foo cover setup for every supported client.