Rennet is a review harness. It helps you work with coding agents effectively and transparently by breaking large changes into something you can understand and inspect.
Not for vibe coders. For agentic engineers.
Available now · Free · Open source
No API-key setup. No telemetry. Uses the Claude Code and Codex already on your machine.
macOS on Apple silicon, signed and notarized. Windows installer, not yet signed, so SmartScreen warns once. Both update themselves.
Pull requests have ballooned. Reading file by file asks you to keep pace with a machine that never gets tired.
Rennet does the structural reading for you: it groups the change, puts it in order, explains it, and surfaces the decisions inside it. You still read the code. Your attention goes where your judgment is actually needed.
Whether the code is yours or a teammate’s, the job is to understand what carries your name.
Your coding agent finished the branch. Before you open the PR, Rennet helps you understand what it did, catch what is wrong, and send the fixes back to the agent, so the review a teammate sees is one you already stand behind.
Review the committed and working-tree changes.
Turn what you find into focused work for your coding agent.
Rennet snapshots the branch again and re-reviews only what changed.
A teammate sends an agent-written PR. Rennet walks you through it in an order you can actually follow, keeps the real code under your hands, and turns your decisions into one normal GitHub review in your own voice.
Read the promise, structure, and changed code together.
Question any finding, decision, or line.
Edit and post the exact review GitHub receives.
Open a change and Rennet reads it five ways. Each lens is a board that answers one question about the branch, and every answer cites the lines it came from. Open a citation and the code appears right under the sentence that cites it.
The screenshots below are the real app reviewing an example branch: per-organisation rate limiting in a small service called atlas.
Design reads the branch against whatever it was written from: a spec, an ADR, a plan, or the pull request description. On atlas it finds the rate-limiting spec and lays out each requirement with the code and tests that answer it. Open one and the cited lines appear under the requirement, so the promise and the code that keeps it sit together on one screen.
Sequence turns the change into steps, each a short explanation over the code it describes. Step two here is the token bucket: take refills from elapsed time on every read, which is why there is no timer anywhere in the branch. The note under the code names the clamps that keep it quiet. You still read the code; Sequence tells you what to read next.
Decisions lists the choices the branch made, each with its reasoning, the alternatives it passed over, and the lines that show it. Five here: two stated in the proposal, three inferred from the code and marked as such. The first is one refill-on-read bucket per organisation, chosen over a fixed window that lets twice the limit through at the edge. Rennet surfaces the call. Whether it was the right one is yours to decide.
Flagged collects what Claude and Codex found, each finding with its severity, the lines involved, and a proposed fix. Three open on this branch, one high: a Redis outage that silently removes every limit. Both models raised it and disagree on how serious it is, so the board shows the split instead of averaging it away. A finding only one model raised says so. Dismiss it, discuss it, or request the change. Rennet flags; the verdict is yours.
Every citation opens the same built-in Diff view: the changed-files rail, the hunk in place with syntax highlighting, and a marker on any line you have a request against. Here it is open on the middleware, with a request sitting on the 429 lines. Move from overview to evidence without losing your place. Rennet changes the reading order, not the need to read.
Ask what changed, why it changed, how it fits the rest of the repo, or whether a failure path is intentional. The conversation stays attached to the review and its code.
− return store.get(key)+ return store.get(key).catch(() => fallback)Why does this fail open instead of blocking the request?
The PR treats rate limiting as protective infrastructure, not an availability dependency. The fallback preserves service when the store is unavailable.
Rennet finds the Claude Code and Codex already installed on your machine and signs in as you already do. No new API keys. No extra bill from Rennet.
Claude and Codex review the same evidence independently. Rennet shows where they agree and, more importantly, where they split. Nothing is averaged into false consensus.
Fallback masks a real storage outage.
Request changeAvailability policy supports this fallback.
Accept with noteModels disagree · human judgment required
Illustrative dual-review outputHighlight any lines on a board or in the diff and the toolbar offers Comment, Request Changes, and Explain. Explain asks the conversation beside the review about exactly those lines. Request Changes records a request against them that your coding agent will receive, with the lines attached. Here the reviewer has selected the failOpen wrapper under the high finding.
Every change you request collects in one place with its cited lines and its thread. Dispatch Round hands the whole set to your own coding agent. When the round is back, Rennet re-reads the branch and shows you what moved: the sections the round touched open, the ones it left alone stay folded, and every request is checked against the round's actual diff, not the agent's word. When nothing is left to ask, push the branch and open the pull request, in your own words and under your own name.
Run Rennet as the desktop app or open it as a browser tab served from your own machine. Both are the full product: no read-only mode, no feature in only one. Pair another computer or your phone over Tailscale to review from wherever you are. Nothing is hosted. The remote device talks straight to your machine, and it never sees a path on it.
Nothing goes to Rennet. There is no Rennet backend and no telemetry. Your coding agents still send context to their own providers, under the accounts you already have, exactly as they do when you use them directly. Rennet says so plainly instead of hiding it.
No. Rennet does the structural reading. You inspect, question, edit, and post the review another person sees.
No. Claude and Codex run independently. Agreement is evidence; disagreement is a signal for you, and Rennet never averages it away.
Rennet itself sends nothing anywhere. The coding agent you use still talks to its provider under its own terms and sign-in, exactly as it does today.
Not for vibe coders.
For agentic engineers.
macOS on Apple silicon, signed and notarized. Windows installer, not yet signed. Both update themselves.