Design partner programme — closed beta

Run more agents
than you can read.

collabd gives one answer on every pull request — safe to ship, needs your review, or do not ship — whoever or whatever wrote it, bound to the exact commit it judged.

1Install 2Pick a repo That’s the setup. It opens pull requests. Nothing merges until you click Ship.
APP-612·collabd-app claude-opus-5
Rate-limit the /events SSE endpoint per session
reviewed c5835e9c1b77

Three answers. Never a score.

Safe to shipBounded, covered, nothing sensitive moved.
Needs your reviewNo risk found, but something wants a human.
Do not shipThe tests pass. It still isn’t safe.

There is no fourth state. When the reviewer cannot reach a judgement at all you get a slate and the reason — never a colour collabd has not earned.

A second pair of eyes that never gets tired of your diffs.

The reviewer is a separate context from whatever wrote the code, so it has nothing invested in the answer.

One answer, not a score

Three states and only three. No percentage, no risk number, nothing to interpret.

Reasons on the line

Every finding is pinned to the exact file and line that produced it, worst first.

You define “done” up front

Approve a plan with acceptance criteria and the verdict grades the diff against them — on the work collabd authors.

Verdicts expire

Bound to the exact commit. Move the code and it goes stale rather than lying to you.

Evidence you can keep

Export the immutable, secret-redacted bundle behind any verdict. It reads fine without us.

Every decision is signed

Who approved, who overrode, who shipped, and what they wrote. Overriding is allowed, and recorded.

You are in the loop twice, and only twice.

Once before anything is written, once before anything ships. Everything between those two moments is the agent’s problem, not yours.

you Describe the work

One sentence. Or let collabd read the repo and propose something itself.

you Approve the plan

Steps, files, open questions. Nothing is written until you say yes.

agent It builds and opens a PR

Sandboxed, on a branch, with its own tests. Never on your default branch.

you Ship what’s safe

The verdict lands and you decide. collabd never merges on its own judgement.

It watches your teammates’ pull requests too.

Same reviewer, same three answers, on every pull request your team opens — whoever or whatever wrote it.

collabdDo not ship
Tests pass, and the change is still not safe: user-controlled input reaches a shell.
Details
ci / build
Successful in 1m 42s — your CI, which collabd does not read.
Details
  • No write access. A separate app that reads the diff and publishes a check. That is the entire surface.
  • Advisory by default. It never blocks a merge unless you make the check required yourself.
  • Its own tests, not yours. collabd runs a sandboxed suite and says so — it never borrows your CI’s word for it.

Least privilege, or it would not be worth installing.

Two scoped Apps

One authors on a branch and opens pull requests. One only reads. Neither merges on its own judgement — a human clicks Ship.

No secrets in the sandbox

Builds and reviews cannot reach your credentials or your infrastructure.

Redacted on the way out

Check summaries, evidence bundles and the audit trail are scrubbed before they leave.

Point it at one repository.

Closed beta, with people already running several agents and losing track of what is safe to merge.