Agent loops that run on their own, within checks you set.
A software factory is many agent loops working at once with nobody driving each step. ProPR runs those loops on your own server: a review and fix loop that works a pull request toward your score, a failing check that starts its own repair, an epic queue that merges one pull request before starting the next, goals that keep going across many steps, and automations that find their own work. Each loop ends at a check you chose, and by default a person merges every pull request. You set the intent, the checks and how far each loop may go. ProPR is free Apache 2.0 software on a server you run.
What makes a factory
Loops that keep working without a person: review and fix, CI repair, conflict resolution, epic queues, goals and scheduled automations.
A check on every loop: CI, the review score, protected paths, cycle and spend limits.
Isolated runs in parallel, with several loops on one repository at once.
You own the outer loop: the intent, the checks and the autonomy level per repository.
Telemetry on every run: prompt, diff, logs, model and cost.
Overview
Underneath every loop runs the same line: plan, isolated run, pull request, review, merge. The workflow page follows one change along it, and the status on each pull request tells you where the loop stands.
The loops ProPR runs.
Each loop has something that drives it and something that stops it. The driver is an event in your repository; the stop is a check you chose. A person reads the result when the loop ends, or earlier when it needs a decision.
Loops around your checksA person merges by default
Youintent, checks, autonomy per repository, the merge
Each loop runs without a person between its steps and stops at the check it names. Review and fix loops work toward Ready within your cycle limit, a pull request that stops short reads Needs you, and you hold the outer loop: what to build, which checks count, and how much autonomy each repository gets.
Swipe for the rest →
Review and fix loop
Ultrafix reviews the pull request, fixes the merge blockers, waits for CI on the new head and reviews again. Every pass lands as a commit and a scored review in the PR thread.
Drives it
A /ultrafix comment, or Run ultrafix after PR on a task or plan.
Stops it
A review score at your goal, the cycle limit, CI that stays red past the wait timeout, or removing the ultrafix label.
Swipe for the rest →
CI repair loop
With Auto CI follow-up on for a repository, a failing check on the pull request head posts one comment naming the check, the commit and the failure output, and that comment starts a fix run. The CI failure playbook walks through one.
Drives it
A failing check run or commit status on the current head.
Stops it
Off by default. Each failing check starts one fix per commit, and checks you mark non-blocking start none.
Swipe for the rest →
Conflict and up-to-date loop
When the base branch moves and the pull request conflicts, ProPR merges the base into the branch and an agent resolves the conflicts. With auto_merge.require_up_to_date, a green pull request behind its base is updated first and merges only after CI passes on the updated head.
Drives it
A push to the base branch, often another pull request merging.
Stops it
Both options are off by default. A conflict the agent cannot resolve reads Needs you, and CI on the updated head decides the merge.
Swipe for the rest →
Epic queue
An epic plan runs its issues as a chain. Each child pull request merges into the epic branch before the next issue starts, and one progress comment on the epic PR counts the merges. The epic playbook shows the whole chain.
Drives it
The previous pull request merging.
Stops it
The auto-merge policy: a pull request that touches a protected path waits for a person, and the queue resumes when you merge it. A blocked issue can start an automation. The chain ends in one epic PR you merge.
Swipe for the rest →
Goals
A goal keeps one agent session working toward a continuing objective across many steps. It commits checkpoints to a draft pull request and asks you only when it needs a decision. Claude Code and Codex run goals natively; Antigravity CLI runs them through resumable sessions.
Drives it
The objective, plus the corrections you send while it runs.
Stops it
A question or approval request waits under Needs you, Pause and Cancel apply at the next safe boundary, and the goal ends in a draft pull request you review.
Swipe for the rest →
Automations
A saved prompt runs on a schedule, or when an epic issue blocks, investigates your repositories and writes a report. In preview or auto mode a second run acts on that report through ProPR's tools: tasks, plans, goals, todos, PR comments, labels, reviews and fix loops. Automations are the loop that finds its own work.
Drives it
A UTC cron schedule, the Plan issue blocked event, Run now, HTTP, MCP or the CLI.
Stops it
Dry run by default, so a run only reports. The capabilities you grant, approval in preview mode, a separate opt-in for merging, and at most three runs per blocked issue.
Outside drivers
The MCP server and propr-cli expose the operations the web UI has: create tasks, plans and goals, start Ultrafix, trigger automations, read a pull request's status. Another agent, a chat client, a CI job or a script becomes the outer loop that drives ProPR.
Drives it
Whatever you connect: a coding agent, a chat session, a scheduler or a pipeline.
Stops it
MCP is off until an administrator enables it. Every call runs inside the grant's scopes and repositories and the user's permissions, and the MCP log records each one.
Swipe for the rest →One automation's runs, started by its schedule, MCP, the CLI from GitHub Actions, the HTTP API and Run now.
What caps each loop.
A loop gets as much autonomy as its check deserves. Verification sets the pace of the factory, so every loop above ends at one of these checks, and each is something you can read in GitHub or on the task.
CI as the merge gate
Nothing merges on a failing check. Ultrafix waits for green before each review, and auto-merge waits for CI on the head it merges. Checks you list as non-blocking keep running without holding a loop.
Scored reviews and the rating goal
Every /review ends with a score out of 10 and findings with fixed IDs. Ultrafix works toward the rating goal you set, and a pull request is Ready only when its latest score meets that goal.
One status per pull request
Working, Needs you or Ready, at the top of the pull request and in its title. Ready means nothing running, every check passed, no conflict with the base and the score at your goal. Needs you means a loop stopped short.
Protected paths and branch protection
auto_merge.protected_paths in .propr/workflow.yml keeps auto-merge off any pull request that touches the globs you list, and .propr/** is always protected. GitHub's branch protection applies on top.
Cycle and spend limits
Ultrafix has a maximum cycle count and a CI wait timeout. A per-run spend cap stops an implementation, follow-up, fix, review or Ultrafix cycle at the estimated cost you set, and the task timeline records why.
Automation capabilities and approval
An automation starts with read access only. You grant web access and ProPR tools separately, preview mode holds the acting step until you approve, and merging needs its own capability.
Restricted network mode
Optional per instance or repository: the agent container starts with no network and reaches only an allowlist of hosts through a proxy on the worker. Denied hosts appear on the task timeline. Read more under trust.
Many loops at once.
Every task runs in its own Git worktree, branch and Docker container, so loops never share a checkout. Add several llm-* labels to one issue and ProPR opens one branch and pull request per model, so you compare implementations of the same change. On one repository an Ultrafix loop works one pull request while a CI repair fixes another, the epic queue advances and an automation writes its report.
Swipe for the rest →Five runs at once, each with its step and elapsed time.
Telemetry on every loop.
Every run keeps its prompt, diff, logs, model, tokens, cost and duration on the dashboard and CLI, and the timeline names the check that stopped it. You see which loop is slow, expensive or wrong. The observability guide lists every record.
Swipe for the rest →The LLM log ties each call’s tokens and cost to the task, plan or repository that spent them.
You own the factory.
Every loop is Apache 2.0 software on a Docker host you control, running Claude Code, Codex, Antigravity CLI, OpenCode or Mistral Vibe on your subscriptions or API keys with no token markup. Compare that with hosted agent-workforce platforms.
Autonomy is a dial.
ProPR is a light factory by default: every pull request waits for a person to merge. You turn the dial per repository. Green pull requests auto-merge under the policy in .propr/workflow.yml, an epic queue merges each child into its epic branch, and an automation's acting step merges ready pull requests once you turn on its merge capability. Every merge still needs passing checks and satisfied review and branch-protection rules, and the acting step never edits or pushes code itself.
A one-person factory, running in public since May 2025.
ProPR builds itself with these loops: 2,400+ pull requests opened, 2,100+ merged, 690+ of them building ProPR itself. Every review score and merge is public.