steward.foo

Question-based PR review for your repositories

Define yes/no review questions. The platform scores every pull request and push against them, surfacing concerns before code lands.

Questions are explicit, testable, and tunable. You see what they catch on historical PRs before they go live. Thresholds adjust based on evidence, not gut feel.

Install the GitHub App

How review questions work

A review question is a yes/no question scoped to a repository. Questions are grouped by the context they need to answer correctly — which files to read, how much budget to spend.

Each question has a threshold that determines when to flag a concern. When the model's probability exceeds the threshold, the concern appears on the PR.

Questions are backtested against historical PRs before going live. You see false positives and missed issues upfront. Thresholds tune against real data.

What gets reviewed

Pull requests

Each PR is scored against your active review questions. Concerns appear as PR comments. A check run summarizes the review verdict.

Pushes to main

Direct commits to the default branch are also scored. Concerns surface in weekly digests instead of on the push itself.

Install

Step 1: Install the GitHub App

A repository admin clicks Install once. This grants read access to code and PRs, and write access to post review comments.

Install GitHub App

Step 2: Connect your coding agent (optional)

Connect your coding agent to define and manage review questions:

https://mcp.steward.foo

Add this URL in your agent's MCP server settings.
Install the Steward GitHub App if you haven't already.
Complete the short-code authentication to connect your agent to the installation.
Specify which repository to target — agents running in a project checkout use that repository by default, or can specify owner/repo in tool calls.

Once connected, your agent can propose review questions, backtest them against historical PRs, and help you tune thresholds.

Full install guide →