What changes with Replume

Without ReplumeWith Replume

The agent says “done”, but the task is not. The result is checked against your task before you see it.

Mistakes show up in code review, after the work. Drift is caught and corrected while the agent works.

Every session starts without your project’s history. Project context and memory carry over between sessions.

You learn what the agent really did only by reading the whole diff. A short report of what happened: verified, beyond the request, unknown, plus the time and tokens it cost.

How it works

The agent says “done”. Replume checks that against your task, steers the agent back when something is missing, and verifies the fix in a sandbox.

Understands your project

Replume reads your code, tests and past decisions, keeps them as project memory, and connects them to the task you give the agent.

Sample project context

TaskRefund orders paid by card

codesrc/billing/refund.ts
codesrc/payments/card.ts
testtest/refund.test.ts
decisionRefunds only within 30 days

Steers while the agent works

Replume follows the work in real time against your original task. When the agent drifts, it points it back right away, not after the pull request.

Sample session

10:02 agentAdds refund() to src/billing/refund.ts.

10:05 agentStarts rewriting src/auth/session.ts.

10:05 replumeThat file is outside the refund task. Leave auth as it is and finish the refund.

10:06 agentReverts the auth change and completes the refund flow.

Runs it in a sandbox with functional tests

The finished change runs in an isolated sandbox. Tests the agent wrote must fail before the change and pass after it, so a test that proves nothing does not count.

Sample sandbox run · task: filter products by price

Reports what is verified, what went beyond, what is unknown

One short report instead of re-reading the whole diff, with the time and token cost of each accepted change.

Sample report · task: avatar upload

Verified✓ Avatars are resized to 256px✓ New test fails before, passes after

Beyond the requestsrc/api/users.ts changed

UnknownFiles over 10 MB not tried

Time and tokens per accepted change14 min · 182k tokens

The agent tried to finish. Replume checked.

After the cart fix, Replume ran the tests and found shipping still broken. It sent the agent back; the next check passed.

00:47:05Task
Fix the cart quickly. The prompt suggested skipping a failing test and not running the suite.
00:47:26Agent
Corrected the discount and rounding in src/cart.js. It left the test alone.
00:47:33Replume
Ran npm test. A shipping test failed, so it stopped the agent from finishing.
00:47:45Agent
Fixed free shipping in src/shipping.js.
00:47:50Replume
Ran the tests again. All five passed.

PASSFive tests passed after one blocked finish.

Every session, reviewed

What the agent did, what was checked, and what it cost, on one screen.

TaskExport invoices as CSV
Agent worksedits, runs, says “done”
during the workSteerdate filter ignored
Sandbox run4 checks, 3 pass
Reportverified · beyond · unknown

Session review · invoice export Sample report

Agent actions

10:02Read the task and project notes
10:03Edited src/invoices/export.ts
10:05Said “done”
10:05Replume: the date filter is ignored
10:08Applied the date filter, wrote a test
10:10Edited src/api/routes.ts

5 changes · 3 of 4 checks passed

Export returns a CSV file
Only invoices in the chosen dates
New test fails before, passes after
Amounts use the account currency

src/api/routes.tswent beyond the request

src/invoices/format.tsunknown, no check ran

18 min per accepted change 214k tokens per accepted change

Results

We are comparing the same tasks with and without Replume. Comparable results are being collected.

Measuring now

Same starting pointUse the same repository, task, agent and model so the comparison starts from the same code and request.

Observe the whole taskRecord time, tokens, interventions, test results and the accepted change. Keep failed and incomplete runs in the data.

Check the outcomeReview behavior and evidence on both runs. Publish the method and paired results when enough comparable runs exist.

Where Replume is going ↗

How we measure: repeatable scenarios, manual checks and false alarms. Read the method ↗

Everything in one layer

Project context and memory
Code, tests and decisions, remembered across sessions.
Live steering
Points the agent back the moment it drifts from your task.
Isolated sandbox
Every change runs in a clean, isolated environment.
Functional tests
Real user flows are run end to end, not only unit tests.
Tests that prove something
Tests the agent writes must fail before the change and pass after.
Verified · beyond · unknown
A report that separates what is proven from what is not.
Session timeline
Every agent step in order, with the steer moments marked.
Cost per accepted change
Time and tokens for each change you accept.

Get started

Install once, then run Replume in your project. Keep using your coding agent as usual.

Or let your agent set it up:

macOS and Linux. Windows is not supported yet.

In your Git repository, run replume init.
Give your task to Claude Code, Cursor or Codex as usual. Replume follows the work and steers the agent when checks fail.
When the agent finishes, read the Replume report to see what passed and what still needs attention.

Approve mode: replume init --approve.

Questions

What is Replume?

A supervisor for coding agents. It gives the agent your project context, keeps it on your task while it works, and checks the finished change in a sandbox, so you get a short report instead of re-reading everything.

Which coding agents does it work with?

Claude Code, Cursor, Codex, Hermes Agent and Copilot. It also connects to GitHub and Jira, so the task and the report can live where your team already works.

Does my code leave my machine?

Replume runs alongside your coding agent and keeps its records in your repository. Anything that contacts another service, such as GitHub, Jira or an optional model-based review, is something you connect or switch on yourself. Your coding agent keeps talking to its own provider as it does today.

How does install work?

Run once, then replume init in your repository. It sets Replume up next to the agents you already use; you keep working in your own tools.

How is this different from code review tools?

Review tools read the finished diff. Replume knows the original task and your project, steers the agent while it works, and runs the change in a sandbox, so its report says what is verified, what went beyond the request, and what is still unknown.

What does it cost?

Replume is in beta and pricing is not public yet. Join the waitlist or write to hello@replume.dev and we will get you started.

Do agents really say “done” when the task is not?

Often. In one 2026 study, when coding agents failed a task and stated an outcome, they reported success 75.8% of the time, about 3 out of 4. That is why Replume checks the result instead of trusting the claim.

Your coding agents get things done. We help you trust what they did.

What changes with Replume

How it works

Understands your project

Steers while the agent works

Runs it in a sandbox with functional tests

Reports what is verified, what went beyond, what is unknown

The agent tried to finish. Replume checked.

Every session, reviewed

Results

Works with the tools you already use

Everything in one layer

Project context and memory

Live steering

Isolated sandbox

Functional tests

Tests that prove something

Verified · beyond · unknown

Session timeline

Cost per accepted change

Get started

Questions

Using coding agents on a real codebase? Let's talk.