Quickstart.

From a new account to the first work Perfloop runs on your code, and where its result shows. One task per step, with the page and the button the app shows.

Steps

1 · get an account

Get an account at perfloop.ai. If a teammate invited you to their workspace, accept the invitation.

2 · open your workspace

A teammate's invitation signs you in to its workspace. Perfloop then opens Start, which lists the setup steps that remain. An Admin takes step 3.

3 · connect GitHub and choose repos

Before you connect, Security lists what Perfloop reads, what it keeps, and what it cannot do. In Setup, select Install Perfloop GitHub App. On GitHub, choose Only select repositories and pick the repos Perfloop may read. Back in Setup, select the repos for Perfloop to work on, then Save selection. Perfloop reads and works only on the repos you choose. Connect GitHub lists the access the App asks for.

Setup, before a code host is connected.
Setup, after the App is installed: choose repos, then Save selection.

A model build then reads each repo, and the hypotheses worth testing appear on the Cases page. The Model page shows the build. If none starts after you save, select Build this repo on the Model page, or ask your coding agent to start one with buildModel. A build is paid work.

what to expect

Many attempts fail the proof; that is expected, and the failed ones stay on record. Start with one Case before you turn on automatic work.

4 · connect your coding agent

Add the Perfloop MCP server to your coding agent and sign in with your Perfloop account. The steps for each agent are on The MCP server, and in Setup under Coding agents.

Setup, Coding agents: connect to https://app.perfloop.ai/mcp and follow the steps for your agent.

5 · start the first work

In Setup under Automatic work, each repo's Case work control, Keep working, starts off. While it is off, Perfloop picks no Case on its own. Start the first work in one of two ways:

  • Hand Perfloop a branch or a pull request. Ask your agent to submit it. The agent calls submitDiff, then reads the findings with diff. A finding is a hypothesis Perfloop has not yet proven. Ask the agent to prove the one you want; it calls workCase. For a pull request, it first asks you to approve running that exact head's build and test code.
  • Start a Case. Open Cases, open a Case in state PROPOSED, and select Work this case. When its case session starts, the Case changes to OPEN. Your coding agent can start a PROPOSED Case too: it calls workCase with the Case's caseID, or with a hypothesisID. A Case found only in a branch or pull request also needs that diff, as above.
The header of a PROPOSED Case: Work this case starts its first case session.

Each is paid work and draws on Usage. For Perfloop to pick and work Cases itself, an Admin selects Keep working for the repo and Save settings. Autonomy and limits explains each control.

New work needs a positive Usage balance. Work already in progress can cost more than the balance left. Unpaid Usage carries into the next Usage period. New funds first cover that shortfall.

6 · read the results

A Case page shows its state and the evidence behind it. VERIFIED means a change passed its checks, every claim held, and the verifier found nothing. Case states explains every state and what to do next.

A verified change can become a pull request, or only an issue when an issue is the better contribution, such as a finding the maintainers should fix themselves. The Pull requests control starts off, so the proposal waits in Inbox: select Approve & open PR (or Approve & open issue) or Reject. Perfloop opens it on GitHub. You review and merge a pull request as usual; an issue has nothing to merge. Perfloop cannot approve, merge, or deploy.

Inbox: a proven change waits for Approve & open PR, or Reject.

Your agent reads the same results with cases, and a submitted change's findings with diff.

Get help

In the app, select Send feedback in the side rail (in the menu on a phone), or write to hello@perfloop.ai. Both reach the Perfloop team.

Also in this section: The loop, Case states, Steering and learning, Initiatives, Autonomy and limits, Members, and The MCP server.