Launching today

Replume
Know what your coding agents actually did
1 follower
Know what your coding agents actually did
1 follower
Ship agent-written code you can trust. Replume supervises Claude Code, Codex and Cursor from task to merge: a contract from your task, live checks on every step, tests judged by content, its own test run, and one verdict with the evidence.









Free
Launch Team / Built With

AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted



Hi Product Hunt, I'm Rafael, the founder of Replume.
Coding agents write most of my code now. Writing stopped being the hard part. Trusting the result is. Again and again I'd find the same things at review: the agent edited files the task never touched, changed a test so it would pass, touched config it shouldn't have, or said "done" when it wasn't. Every one of those costs another prompt, another wait and another rerun, and some of them I only caught after merging.
So I built Replume: a supervisor that sits next to your coding agent and watches every step, from the task to the merge.
What it does today:
· Understands the task. It turns your prompt and repo into a contract: what's in scope, which tests the task actually asks to change, what has to stay as is. It records where everything stood when the task started.
· Watches every step live. Every edit and command is checked against that contract as it happens. When the agent drifts off the task, it gets steered back right away, and anything beyond the request is listed for you.
· Keeps the tests honest. Test changes are judged by what they actually change: an expected value that moved (old → new), a dropped case, a weakened or hidden assertion, a test quietly skipped in config. The agent is told immediately, and a changed test can't end in PASS until a person confirms it.
· Checks the work itself. Replume reruns your tests on its own instead of trusting the agent's output, and checks the agent's claims against what really happened. If the agent admits something is unfinished, you get INCOMPLETE, not PASS.
· Tells you where you stand. One verdict (PASS, FAIL or INCOMPLETE) with the reason and the next step. replume watch shows the session live, replume stats shows what it caught and the time it saved you, and replume share makes a card of the session.
· Follows the work to the PR. replume init --ci adds a GitHub check that puts the session's evidence on the pull request, and replume verify --task checks work done without an agent session.
It works with Claude Code, Codex and Cursor, runs locally, and your code never leaves your machine. Free: npm i -g replume-cli, then replume init. macOS and Linux, Windows is experimental.
Where it's going: Replume learns each repo (which tests really run, what was already broken, where agents cut corners before), so checks get faster and sharper over time. Next up: a verdict you can review in a minute (fix first / review / safe to ship, each with its evidence), PR checks that count a reviewer's approval as the human confirmation, an MCP server so any agent can check itself before saying done, and replaying real requests against your app to show a change actually works. For teams: shared rules, history and reliability per agent.
I'd love to hear where it got in your way or what it missed. That's exactly what I'm working on.