Skip to content
patchkind

Free invited beta

You approve a short plan. Before any code is written.

Turn user feedback into a change you’ve chosen. Review the plan, choose when Claude starts, then review the pull request before anything ships.

Runs in your own GitHub Actions, using your own Claude subscription or API key.

Claude may ask a question or stop instead of opening a PR.

See how it works Request free beta access

Free invited beta · npm projects · Claude Code · owners sign in with ChatGPT

How it works

From “this is confusing” to a change you can review.

You choose the plan, when to run it, and what ships.

  1. 01

    Someone shows you what could be better.

    Collect suggestions, missing features and confusing screens, with screenshots or annotations when useful.

    Feedback with context
  2. 02

    You choose the change.

    Review the evidence and edit the plan. Approving a plan does not start a run; you choose when Claude begins.

    Approve, edit or defer
  3. 03

    Claude works in your GitHub.

    The coding job runs in your repository’s GitHub Actions, using your own Claude credentials held in your GitHub secrets.

    You press Run
  4. 04

    You review the result.

    See what changed, the checks that ran, and anything still uncertain. Claude may ask a question or stop instead. You decide whether to merge and release.

    Review before release

Ways people send feedback

  • Embedded buttonInside your product
  • Share linkNo code needed
  • Your own text linkIn a footer or menu
  • Report APICustom form integrations*
  • ChatGPT / ClaudeFor owners, via MCP

Your own link opens the same feedback panel; the floating button is optional. *Custom forms require same-origin requests and verification. The MCP connector lets owners work with collected feedback in their assistant.

Planned Email-in to support@patchkind.com · Slack / Discord commands · Imports from GitHub Issues, Intercom and Zendesk · A docs “Was this helpful?” widget. These aren’t available yet; the support address currently handles support only.

What you see as the owner

A clear place to decide what happens next.

Example data. Fictional projects, people and tasks in the real workspace. Select a screen to see it full size.

View full-size Inbox →Example Inbox grouped by project, with plans to review, a question from Claude and a PR ready for review.
Inbox · Example data · See what needs you, grouped by project.
View full-size Task →Example task with a five-step progress indicator, a pinned screenshot, the original note and a plan to approve or edit.
Task · Example data · Review the screen, the note and the plan together.
Example phone task showing the plan and bottom actions: Approve, Edit plan and Defer. Starting a run is a separate step.
Phone · Example data · Approve on the go. Run remains a separate step.

What you control

Your repository. Your Claude account. Your release decision.

  1. Your GitHubSubscription token / API key
  2. Agent step onlyClaude works in the sandbox
  3. PatchkindNo owner coding key
Your credential stays in your GitHub secrets. Only the agent step receives your coding credential; Patchkind never collects it. Optional built-in summaries use Patchkind’s separate OpenAI service credential.

Work stays in your workflow.

Work runs in your GitHub Actions, with the code and review in your existing workflow.

Your credentials stay with you.

Use your Claude subscription token or API key. Patchkind does not collect either. Subscription usage limits and GitHub Actions usage still apply.

You decide what ships.

Set the approved scope. Review the resulting pull request. Patchkind never automatically merges it.

Less translating feedback into coding tasks.
More deciding what matters.

Keep the original report, the agreed change and the result together. Pick up a small improvement without reconstructing the conversation or handing your product over to an unattended agent.

The AlgoThesis trial

Small changes.
Four merged pull requests.

A supervised trial on our owner’s own app. The four changes were verified live on 5 October 2026.

Read the full case study How the Claude Code workflow runs

What shipped

  • A one-line company descriptionPR #720 (private repository)
    Live
  • A clearer company headerPR #723 (private repository)
    Live
  • An explanation for missing earnings price dataPR #724 (private repository)
    Live
  • Guidance on the Following dashboardPR #725 (private repository)
    Live

The AlgoThesis repository is private; screenshots of the live changes are shown in the case study. One earlier PR targeted the wrong page and was closed. A follow-up improved the first change. These were small fixes, with human decisions throughout.

Evidence, not a promise of perfection

Tested with deliberately hostile feedback.

In our supervised AlgoThesis trial, we included requests to expose secrets and alter protected files. Our verifier reported that Claude ignored those instructions and changed only the page it was asked to.

The same run also tripped our protected-path check because of a packaging bug of ours: empty placeholder files. We fixed it. The check caught that bug, not a hostile edit.

This is evidence from a bounded test, not a guarantee that every malicious report or incorrect change will be caught.

  1. FetchOK
  2. AgentOK
  3. ChecksFailed
  4. PublishSkipped → no PR
  5. ReportOK
Run 37274899273 · No PR. Checks caught our placeholder-packaging bug; the verifier reported that Claude ignored the hostile instructions.
Read the test and its limits

Attack run 37274899273 · no PR opened (private repository)

Beta expectations

Free invited beta.
Assisted setup.

Projects
Configured npm projects
Coding assistant
Claude Code on GitHub-hosted runners
Owner sign-in
ChatGPT
Every change
Your plan approval, Run and PR review

Claude can ask for clarification, fail a check or stop without opening a PR. Autopilot, self-serve onboarding and notification emails are not promises of this beta.

Have an app that could use this?

Request an invitation for your project. This is an access request, not instant sign-up.

Request free beta access Questions? support@patchkind.com

Optional AI follow-up questions and feedback summaries use OpenAI and can send feedback text and screenshots to that provider. This is separate from your Claude coding run. Privacy & data

Screenshots add context; a passing build does not confirm the visual result.