New Governance for AI-assisted work

Run, watch and govern every AI coding agent.

One web app for the agents your team already uses. Launch Claude Code, Codex or Aider on your own machines (any other terminal tool runs in a shell session), get told the moment one needs you, and keep everyone on the same rules.

Outbound only, no VPNMac, Linux, WindowsNever sends your code
Works withClaude CodeCodexCursorOpenCodeAiderGemini CLIGitHubGitLabBitbucketAzure DevOps
Take the tour

Everything your team needs to put agents to work.

The live terminal of an agent session in Patchkin, paused on a permission question.The Agents page: five sessions across three machines, one marked as maybe waiting for you.Team standards with the tools each applies to, and three machines up to date.Governance findings: a critical secret finding and commits on a protected branch.The Agent tools page: MCP servers across three machines, marked approved or not approved.Usage and cost: cost per day, by model and by tool.
BEFORE

Work is invisible

“Who ran that agent, and what did it change?”

WITH PATCHKIN

Every session, commit and finding in one list.

BEFORE

Agents stall quietly

An agent waits an hour for a yes nobody saw.

WITH PATCHKIN

An alert on your phone, and the question in the app.

BEFORE

Setups drift

Nine tools, nine rules files, every laptop different.

WITH PATCHKIN

One standard, synced everywhere, removable.

BEFORE

No guardrails

Secrets and force pushes found after the fact.

WITH PATCHKIN

Flagged or blocked, with a required pipeline check.

How it connects

Your machines call out. Nothing calls in.

A small program on each machine opens one encrypted connection to Patchkin. Your browser and your pipeline talk to Patchkin too. That is the whole network story.

  • No inbound ports, no VPN, works through an HTTP proxy
  • Your code stays on your machine
  • Optional end-to-end encryption for a machine's terminal

The full picture · Security details

Your machineClaude Code, Codex, Cursor CLI, OpenCode, Aider, Gemini CLI
PatchkinThe web app, your rules and the records
Terminal viewmachine to Patchkin
What the agent prints, streamed to your browser over TLS. With end-to-end encryption on, Patchkin cannot read it.
Small eventsmachine to Patchkin
A session started, a commit finished, a check passed. Names, counts and results only.
Your instructionsPatchkin to machine
Launch this agent, send this input, apply this standard, use this policy.
Never crossesstays on your machine
File contents, what people type, secret values, the output of your tests.
USAGE AND COST

What agents cost, by tool, person and repository.

Counts only. Cost per merged pull request, using your git host.

Usage and cost in Patchkin.
AI ATTRIBUTION
3 of 8

commits in this demo were AI-assisted, recognised from commit marks and sessions.

A tool that leaves no mark is only counted when it runs through Patchkin, and the product says so.

THE PART NOBODY CAN SKIP

One required check in your pipeline.

Hooks on a laptop can be skipped. A required check cannot.

# any CI that can run a command
patchkin-daemon verify \
  --hub https://your-patchkin \
  --base origin/main
IDENTITY AND ADMIN

Sign-in you control.

OpenID Connect and SAML single sign-on, SCIM provisioning, two-factor sign-in, roles and teams, and an activity log.

  • Owners, admins and developers
  • Optional end-to-end encryption per machine

“I want my agents to work while I am away from my desk.”

  • Check and steer a run from your phone or another computer
  • Get told the moment an agent needs an answer
  • Carry the work on to a faster machine, with the notes and recent output
  • Use Claude Code today and Codex tomorrow without changing your habits

“I do not want to be watched.”

Fair. Patchkin records names, counts and results. It never records file contents or what you type, and you always see your own activity first.

  • Your prompts and your code are not collected
  • Your organization must show you a notice of what is recorded
  • Per-person views are off unless your organization turns them on

Names and counts, not contents

Governance records branches, commit names, file names and line counts. Never file contents or what people type.

Observe first, enforce later

Start by recording. Nothing is blocked until you decide, and you choose which rules block.

Honest about limits

Hooks on a laptop can be skipped, so the strongest rules go in a required pipeline check. We say so plainly.

Questions

Questions we hear first.

Do I have to change how I work?

No. You keep using your terminal and your tools. Patchkin adds a way to reach them from a browser and a record of what they did.

Is it another coding agent?

No. It does not write code. It runs and governs the agents you already chose, and it works with new ones as they appear.

Where does my code go?

Nowhere new. Agents run on your own machines. Patchkin sees session names, commit and file names and counts, and the results of checks. It does not store file contents or what you type.

Which tools does it support?

You can launch Claude Code, Codex and Aider from the web, and run any other terminal tool, such as Cursor CLI, OpenCode or Gemini CLI, in a shell session. Past conversations can be browsed for Claude Code, Codex, OpenCode, Aider and Cursor (Cursor is view-only). Standards are written for nine tools, and AI attribution recognises the marks those tools leave in commits. We have tested Claude Code end to end; the others follow each tool's documented behavior. See the support table.

Can I try it on one machine first?

Yes. Pair one machine, run an agent, and switch governance on in observe mode, which only records. Nothing is blocked until you choose to enforce.

What happens if Patchkin is down?

Your agents keep running, because the work happens on your machines. The browser view comes back when the connection does.

See it on your own machines.

Pair one machine, start an agent from your browser and watch it work. Tell us about your team and we will set you up.