PRBee
First month free · no card needed

Your pull requests keep moving while you do something else.

Your AI agent, Claude Code or Codex, reviews your PRs, answers threads and fixes red builds and conflicts. On your own machine, within the hours and budget you set.

curl -fsSL https://prbee.dev/install.sh | bash

Paste it into a terminal, on a Mac or on Linux (Ubuntu 22.04+, Debian 12+, Fedora 36+; x64 or arm64).

irm https://prbee.dev/install.ps1 | iex

On Windows 10 or 11, paste this into PowerShell instead; nothing needs administrator rights. Or download the zip and run its install.ps1.

On a Mac you can also download the .dmg, for Apple silicon or Intel (not notarized yet: allow it once under System Settings > Privacy & Security). All releases

  • Runs only on your machine
  • No account, no telemetry
  • Nothing is ever sent to us

PRBee in a minute

What it does

It finds the pull requests waiting on you, decides what each one needs, and works through them one at a time: one agent session per job, with its cost, tokens and log kept for every run.

Reviews

Reviews your colleagues' PRs

Reads the diff, and as far around it as the depth you choose, then posts findings on the lines they are about.

Reviews your own PRs

The same pass over your drafts before anyone else sees them, so the obvious things are gone before review.

Follows up

After a new push it reads only what changed since its last review, and answers the author in the thread of each finding they replied to.

Clears stale change requests

Once every thread on your PR is answered and the reviewer has stayed quiet for the days you set, it dismisses the old request.

Your pull requests

Answers review threads

Makes the change a reviewer asked for, commits, pushes and replies in each thread. If it cannot decide, it holds the thread for you.

Fixes red builds

Triages the failing CI run first. A failure the branch caused is fixed by re-running only the check that failed; a flaky or shared one is left alone.

Resolves conflicts

Rebases onto the base branch and resolves the hunks, then checks the result before anything is pushed.

Keeps branches current

Brings a branch that is only behind its base up to date through GitHub or GitLab, with no session and no cost.

Merges when it is ready

When a PR is approved, green and has nothing left to do, it can merge it for you. This one is off until you turn it on.

Tidies and labels

Takes the noise out of the comments a PR added, and adds the labels it has earned under your repository's own label names.

Your team

Polite nudges

One short line on a colleague's PR when its build is red or its branch conflicts, so a review does not stall in silence.

Comment rules

Posts the comment your team's workflow is triggered by, once the conditions you set for it hold.

Team view

Review traffic per person: who reviewed what, how quickly, and what reached production.

It gets better the longer it runs

Every review your team does teaches it something. What it learns stays on your machine, and every piece of it can be switched off or deleted.

Lessons from your reviewers

When a colleague catches something PRBee's review missed and the author takes it, PRBee keeps it as a short, general lesson tied to the files it applies to. The next PR that touches those files is reviewed with that lesson in mind, so the same mistake is not missed twice. A report shows which lessons led to findings and which ones authors accepted.

Your repository's rules

It reads your CLAUDE.md, AGENTS.md and .claude/rules files, and hands every job the ones that apply to the paths it changes, the same way for Claude Code and Codex. When those files change, it checks its own instructions against them and adapts. A report shows what each rule cost and what it found.

It investigates production, too

Connect your monitoring and PRBee watches what your last deploy did. Press investigate on an error, an alarm or a slow endpoint, and the agent reads the code and the evidence and tells you the cause in plain words.

New errors

The errors first seen since the last deploy, from Honeybadger, with the users they hit.

Alarms and odd graphs

CloudWatch alarms and dashboards, read hourly, flagging any hour that does not look like the same hour last week.

Slow endpoints

From Skylight: which endpoints are slow, which got slower, and which ones run too many queries.

Evidence, not guesses

While it investigates, the agent can query a read-only replica of your database and check Postmark for the emails involved.

Integrations

A GitHub or GitLab account and one agent are all it needs. The rest are optional, read-only unless noted, and stay off until you add a key.

GitHub

Pull requests, reviews, checks and labels, through the GitHub CLI or a token, with the account you choose.

One forge required

GitLab

Merge requests, discussions, pipelines and labels on gitlab.com or your own GitLab server, through the GitLab CLI or a token. Every job runs the same as on GitHub.

One forge required

Claude Code

Anthropic's CLI, with its own sign-in or an API key. Every cost ceiling PRBee sets is enforced.

One agent required

Codex

OpenAI's CLI, with its own sign-in or an API key. Switch between the two at any time.

One agent required

Trello

Your board at a glance, cards matched to their PRs, and, if you switch it on, your cards moved along as their PRs move.

Writes only if enabled

Honeybadger

Errors introduced by the last deploy, and one-click investigations of their cause.

AWS CloudWatch

Alarms and dashboards, with a read-only key and no AWS CLI needed.

Skylight

Slow and slowing endpoints, with their traces and queries.

Postmark

What was sent, to whom, and whether it arrived, as evidence during an investigation.

Your database

A read-only replica, queried inside read-only transactions that are always rolled back.

Rules it cannot talk its way around

The agent decides what to write. Whether it leaves your machine is decided by fixed rules in the app, not by a prompt.

Checked before every push

Commits are signed, no AI attribution is added to them, and shared CI files are never changed behind your back.

Within your limits

Working hours and US holidays, jobs per hour, and a cost ceiling per job, with today's, this week's and all-time spend on screen.

Knows when to wait

It holds the queue when GitHub is having an outage or your Actions minutes run out, instead of spending money on jobs that cannot work.

Sees trouble early

It warns when two open PRs cannot both land as they are, and double-checks its own reading of CI against GitHub's.

Never in your checkout

It works in its own worktree beside your clone, so your branch and your uncommitted work are never touched.

One switch to stop

Pause everything from the dashboard; a running job stops with it.

Works with your stack

It detects the language of each repository and knows how to install, test and lint it. For anything else, the agent reads your checkout once and learns how.

Ruby on RailsRubyJavaScript / TypeScript PythonGoRustJava KotlinC# / .NETPHPand others

The dashboard speaks English, Turkish and Spanish.

You decide what it does

Every kind of job has its own switch on the Settings tab: reviews, replies, CI fixes, conflicts, branch updates, nudges, labels. Merging and comment rules start switched off, because they cannot be taken back. Nothing runs at all until you finish setup and press Start, and one switch pauses everything.

Getting started

  1. Install PRBee with one Terminal command. It opens in your browser and starts at login.
  2. Pick the agent: Claude Code or Codex, signed in with its own account or an API key.
  3. Let it check git and the CLI for your forge, gh or glab; it installs what is missing. A token works instead.
  4. Point it at a clone of your repository. It works in its own worktree beside it.
  5. Choose the GitHub or GitLab account it acts as, then press Start.

Pricing

The first month is free, with everything switched on and no card needed. After that it is $5 a month, or $50 a year with two months free. See pricing

It runs on your machine. Nothing comes to us.

PRBee has no server, no account and no telemetry. We never see your code, your repositories, your reviews or how you use it.