Perfloop docs.

Guides for Perfloop, and the explanation they assume. The connect guides are written for the engineer doing the setup and for the security reviewer who checks the access it asks for. The work pages are for the engineer who runs Perfloop on a repository and reads its results.

How it works

Start here. It explains the model Perfloop builds of your software, the patterns it hunts on that model, the harness that proves a change, and the direction that stays with you. Its links open one merged case end to end.

Connect your systems

Guides for telemetry sources and code hosts, and private packages. Each guide lists the access to grant and the setup steps, then the data Perfloop reads and stores.

• Telemetry sources. Axiom, Prometheus®, VictoriaMetrics, Pyroscope, Loki, Google Cloud Profiler, Datadog, Polar Signals, Grafana, Grafana Cloud, and Go pprof.
• Code hosts. GitHub, GitHub Enterprise Server, and GitLab.
• Private packages. GitHub Packages NuGet.
• Inside your network. A source the internet cannot reach connects through the Perfloop relay. It reads Prometheus® and VictoriaMetrics for metrics, Loki for logs, and Go pprof for live profiles of Go processes.

Work with Perfloop

These pages assume How it works, except the quickstart, which takes a new account to its first work. The others cover how the loop runs on a repository, what each case state means, how to steer the work and what Perfloop learns per repository from the way you steer it, how an Initiative gives it a direction, what Perfloop may start on its own, and how a coding agent works with it over MCP.

• Quickstart. From a new account to the first work Perfloop runs, one step at a time, and where to read the result.
• The loop. The reference for its objects and results: hypothesis, case, candidate, claim, checks, verdict, and the production check.
• Case states. What each state means, and what happens when work stops or retries: budget, time, replacement, conflicts.
• Pull requests. For the reviewer of a pull request Perfloop opened: how to ask for a change, and what it does by itself.
• Steering and learning. Feedback, scope, closing with a reason, and the lessons Perfloop keeps per repository from each.
• Initiatives. One outcome over part of your system, the research on it, and what its page shows.
• Autonomy and limits. Up to five controls per repository, research, Usage, and the limits that pause work.
• Members. The three roles, the member list, changing a role or removing someone, and single sign-on.
• The MCP server. The agent guide: the first run, the core jobs as tool calls, connecting a coding agent, and every tool.

Questions: hello@perfloop.ai