Linux, macOS, and Windows
Same npm runner on major desktop OSes; use PowerShell or terminal per the npm README for Windows setup.
A local runner that drives Cursor agents through build, then draft PR, then automated review and fix rounds — on your own machine.
npx -y prgenie@latest init
Requires Node.js 22+, git, GitHub CLI (gh), and cursor-agent for
real jobs.
Wire MCP stdio (prgenie-runner) or the optional HTTP dashboard, then start
jobs. Merging the PR stays your call.
Call start_job with a repo path, branch name, and prompt file (via MCP or
the dashboard).
The implementor agent works in a git worktree and lands commits on your feature branch.
The runner commits, pushes, and opens a draft GitHub pull request when the build stage succeeds.
Reviewer agents approve or request changes; fixer agents address findings. The loop runs for up to your configured review rounds (default 3, max 10). You promote or merge when ready.
Same npm runner on major desktop OSes; use PowerShell or terminal per the npm README for Windows setup.
Pick models for build, review, re-review (round 2+), and fix via
stageModels or runner defaults in Setup.
Optional prgenie serve UI on localhost; bind 0.0.0.0 to
reach it from other devices on your network.
Stdio MCP (npx -y prgenie mcp) needs no token; HTTP MCP and the dashboard
use a bearer token from prgenie init.
Jobs queue with a configurable concurrency limit (maxActiveJobs, up to
8).
Runs locally with your own git remotes and gh auth — no hosted agent
farm.
Install once per machine. Server id prgenie-runner keeps the runner separate
from the Cursor plugin's prgenie steward MCP.
npx -y prgenie init
npx -y prgenie doctor
npx -y prgenie serve
Open http://127.0.0.1:9473/ and paste the token from
runner.token. Wire MCP with
npx -y prgenie config --client cursor when you are ready for stdio jobs.
PR Genie by radiancelux is the open-source local runner published on npm
as prgenie — not the vLLM hackathon project, prgenie.co, or other namesakes.
Source:
github.com/radiancelux/prgenie.
Licensed under MIT.