Independent verification for coding agents

Windkeel. A staff engineer for every engineer.

Coding agents write the change. Windkeel checks it the way a staff engineer would before it reaches you: a verifier from a different vendor than the agent, in a fresh sandbox, decides whether it passes, and the pull request arrives with the evidence.

The source is public, so you can run and inspect it yourself, or let us run it for you.

Sign-up is invite-only for now.

Draft

Retry webhook deliveries with backoff

Verdict

pass on attempt 2 of 3

Maker CodexVerifier a different vendor

Acceptance criteria

  • Failed deliveries retry up to 5 timespass
  • Delay doubles between attemptspass
  • A final failure is recorded, not droppedpass

Checks, in a fresh sandbox

  • Repository checks · npm test, lint, typecheckpass
  • Adversarial review · 0 blocking findingspass
  • Security scan · no new issuespass
  • Browser flow · no UI change in this taskskipped
evidence packagesealed with Sigstore
Illustration of a draft pull request and its verdict. Not a real run.

How it works

The maker is the wind. The verifier is the keel.

Wind drives a boat forward; the keel keeps it on course. In Windkeel, a maker agent moves the work, and a verifier it doesn't control decides whether the result holds.

  1. Intake

    1Task

    A title, a description and acceptance criteria, against one repository.

  2. Maker sandbox

    2Maker

    Codex, Claude Code or OpenCode, with the model you choose, implements it.

  3. Verifier sandbox, fresh

    3Verifier

    A different vendor re-runs your checks, reviews adversarially, maps each criterion to evidence, scans for security issues and can drive the app in a browser.

  4. Maker sandbox

    4Repair

    On a fail, the maker gets the findings and repairs, within caps on attempts, spend and time.

  5. Your repository

    5Draft PR

    A draft pull request with an evidence package. A person reviews and merges as themselves.

Steps 3 and 4 repeat until the verifier passes the change or a cap is reached. Each check starts from a clean sandbox and takes nothing from the maker except its commits.

Why an independent verifier

The author shouldn't grade its own work.

  • 01

    A different vendor

    The agent that wrote the code shouldn't be the one grading it. The verifier runs a model from another vendor than the maker.

  • 02

    A fresh sandbox

    Verification starts clean. It takes nothing from the maker except its commits: no files, dependencies or environment.

  • 03

    Evidence anyone can recompute

    The verdict comes with what was checked and why. Anyone can recompute it offline, without asking Windkeel.

The evidence package

A verdict you can check without us.

Every run ends with an evidence package. It can be sealed with Sigstore, so cosign verify-blob alone can check it. You don't need Windkeel to check Windkeel's output.

Three verdicts, never softened

pass
The change meets its criteria and passed every check that ran.
warn
Something needs a person's judgment before merging.
fail
The change does not meet its criteria or a check failed.

What's inside

  • What was checked, and what each check found
  • Transcripts of the review
  • Each acceptance criterion mapped to its evidence
  • Which agent and model wrote the change
  • Checks that did not run, recorded as skipped with a reason
  • The verdict: pass, warn or fail

Re-check it offline

npx windkeel verify-package  # re-check an existing package offline

Why not just run an agent in the cloud

Cloud is where the work runs. It isn't a check.

  • It keeps working when your laptop closes

    Every task runs in a cloud sandbox, not on your machine. Delegated work shouldn't need your laptop awake.

  • But a cloud machine doesn't check the work

    A sandbox isolates the work. It doesn't prove the code is right.

  • And a passing test is still a self-report

    Tests, screenshots and previews the agent produced are useful evidence, but they are the agent reporting on itself. Windkeel adds a check the agent doesn't control.

Start work from anywhere

Work arrives where your team already is.

Every door ends in the same task: a title, a description and acceptance criteria.

  • Web console

    Describe the task and pick the agent.

  • GitHub issue label

    Label an issue to start a run.

  • Slack

    Ask from the channel you're in.

  • Linear

    Hand over an issue.

  • Jira

    Hand over a work item.

  • Signed webhook

    From any system that can POST.

  • CLI and API

    From a script or your CI.

  • Cron schedules

    Recurring work on a timetable.

Two ways to use it

Start with the gate. Add the factory when you want it.

Source-available · free for noncommercial use

Windkeel Gate

A CLI and GitHub Action that verifies pull requests in your own CI.

  • Verifies any pull request, whoever or whatever wrote it
  • Writes the same evidence package
  • Runs in your own CI with your own model keys
  • No servers
- uses: jyothepro/windkeel@<pinned-sha>  with:    mode: verify    anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}

Pin the action to a commit SHA. Your model key stays in your repository secrets.

Managed

Windkeel Cloud

The full factory: intake, maker, verifier, repair loop, draft pull requests and a console.

  • Intake from the console, GitHub, Slack, Linear, Jira, webhooks, the API and schedules
  • Maker, independent verifier and repair loop, with caps on attempts, spend and time
  • Pick the agent and model, and watch the session live
  • Take over in a workspace of your own, then hand back
  • Draft pull requests you merge as yourself

Pricing for Cloud

Pay for compute. AI at cost.

One credit is US$0.01. A sandbox-minute costs 2 credits, for the maker and the verifier alike. Verification is metered like everything else and never skipped to save money.

Models run on your own Anthropic or OpenAI key. Credits pay for sandbox compute only, for the maker and the verifier.

  • Starter

    $20/ month

    2,000 credits a month

    Up to 1,000 sandbox-minutes of compute.

  • Recommended

    Pro

    $50/ month

    5,250 credits a month

    Up to 2,625 sandbox-minutes of compute.

  • Team

    $100/ month

    11,000 credits a month

    Up to 5,500 sandbox-minutes of compute.

Need more? Talk to us.

Billed monthly. Plan credits reset every month.

Top-ups, valid for 12 months: $10 for 1,000, $25 for 2,500, $50 for 5,000 credits.

Overage is off unless an admin turns it on, with a monthly cap.

FAQ

Questions

Merge agent pull requests on evidence, not on faith.

Windkeel Cloud is invite-only for now. Gate is free for noncommercial use today.