Run Claude Code, Codex and other coding agents on your own server. Every task runs in an isolated workspace and lands as a pull request you refine in GitHub.
Installs in two commands via npm
ProPR 0.9.0 · October 2026
Hand an agent a standing objective. Follow its progress, correct it mid-run, and review the result as a draft PR.
Run a saved prompt on a schedule or from CI. Each run writes a report, and you choose whether it becomes tasks.
Switch agents to restricted mode: no network, only a per-run allowlist proxy. Task agents get read-only tokens; ProPR does every push.
ProPR for Linux and macOS: saved instances and accounts, native menus, and guided local setup on Linux.
Run plans, goals, tasks, reviews and automations from an MCP client. An administrator enables it, and you choose which repositories and scopes it gets.
Install the web app, get push notifications per category, and set quiet hours.
Label a GitHub issue and get a pull request back. The agent works on its own. Once the PR is open, steer it with ordinary review comments.
AI labelTriggers the agent run.See the comments, commands, and commits behind three everyday workflows.
Leave a normal GitHub comment. ProPR runs the follow-up and adds the change to the same PR.
Follow the PR feedback loop →Ask a second model to review the code, inspect its findings, and choose which fixes to apply.
See cross-model review in action →Give the agent the failure output, review the correction, and rerun checks on the same branch.
Work through a CI fix →MCP connects chat to the full workflow: discover repositories and models, plan issues, steer goals and tasks, inspect diffs, review and fix PRs, and request a guarded merge.
The optional MCP server caps every chat client at a scope ceiling. Each client connects through browser consent, and you pick the repositories it reaches. Agents and execution remain on your infrastructure. ProPR exchanges text and structured data. Your chat host handles any speech features.
Explore MCP from chat → Set up MCP. Compare access and privacy for each route.
Large features need a plan before code. Finished pull requests need a second opinion from a different agent. ProPR builds in both, issue planning and code review, and returns every task and finding to GitHub.
The planner turns repository context and requirements into review-sized tasks with specifications, acceptance criteria, and a suggested implementation. Refine the plan in conversation, then create the GitHub issues.
Explore planning
Comment /review and ProPR reads the source issue, prior branch changes, current diff, PR conversation, and changed files. It posts numbered merge blockers, optional suggestions, and a score in GitHub without modifying code, then a /fix line that applies exactly the IDs you keep.
Every agent run gets its own Docker container and Git worktree, so parallel agents never collide.
A task-specific branch, Git worktree, and Docker container separate every implementation from the main checkout and concurrent runs.
Each task keeps its prompts, streamed logs, commits, per-call model cost, outcome, and recovery state.
Route the same workflow through Claude Code, Codex, Antigravity CLI, Mistral Vibe, or OpenCode using your existing AI subscription or API key.
Epic mode with auto-merge runs a multi-issue plan in order. Each pull request merges before the next implementation starts.
Direct agents fit live exploration. ProPR fits planned development that needs a visible task history and GitHub review.
Claude Code, Codex, and other CLI sessions fit hot-reloading servers, continuous log tailing, and tight interaction with a live environment.
ProPR runs delegated work in a separate branch, Git worktree, and Docker workspace, then keeps the prompt, diff, logs, cost, follow-ups, and review in GitHub.
Direct and delegated work share one process: debug directly, commit useful progress, then bring the branch into ProPR for planning, follow-up, review, and merge.
Since May 2025, every product we ship has moved through ProPR's own plan, run, and review loop — including ProPR itself.
ProPR is free, open source, and self-hosted. Repositories, task history, logs, and credentials stay on your server, behind a documented security model. Use the AI subscriptions you already pay for, or API keys with metered billing. ProPR never marks up your tokens.
Runs on Linux or macOS with Docker.
npm install -g propr-cli
propr setupProPR is for planned, PR-centered engineering changes. For the deeper operational and security questions, see how ProPR works.
Sign in to your own ProPR instance with GitHub. ProPR is self-hosted. ProPR Connect manages webhook routing, seats and optional managed tunnels. Explore the product at demo.propr.dev; the demo rejects changes to data. Start your own instance with the quickstart.
Your instance serves the HTTP API under /api. Read the API reference and documentation, browse ProPR on GitHub, or join the Discord community.
ProPR is the PR-centered process around coding agents. You start with a plan, run the task in an isolated workspace, get a pull request, and keep refinements inside GitHub. Claude Code, Codex, and other coding agents write the code. ProPR gives them a shared engineering process.
Because that is where teams already review code. The diff, comments, checks, approvals, and merge decision stay in the place engineers trust. Agent output stops living in terminal scrollback or chat history and starts living in the same pull request loop as human-written code.
A terminal session asks you to approve commands and edits as it goes. ProPR delegates that control to the isolated workspace and moves your approval to the pull request. Scope the task well and it works like delegating to another engineer — label the issue, walk away, review the PR. Follow-up is lighter too: screenshots paste straight into GitHub comments, ready for the next follow-up.
Not yet. ProPR is built on GitHub pull requests, review comments, status checks, and comment commands. GitLab and Bitbucket are on the roadmap.
Each task runs away from the main checkout in its own Git worktree, branch, and Docker workspace. Task agents get a read-only GitHub token; ProPR makes every branch, commit, and push itself. Optional restricted network mode starts agent containers with no network and routes their traffic through a per-run allowlist proxy: the model provider, GitHub, package registries, and the hosts you add. Every result returns through commits and a pull request that people can review, retry, discard, merge, or revert. Revert the commit or close the PR — the record of the attempt feeds the next one.
With automatic CI follow-up on, ProPR starts a fix when a check fails, working from the check's output and annotations. The correction lands as a follow-up commit on the same branch. Ultrafix waits for green before it continues. Without it, paste the failing output into a PR comment and ProPR takes it from there.
No. ProPR runs on the provider subscriptions you already pay for, including Claude Pro and ChatGPT Plus. Direct API keys work too, with metered billing. Use one provider or several, and route each phase to the agent that earns it. Either way, ProPR never marks up tokens.
Yes. Add a trigger label (such as AI or propr) to an issue and ProPR picks it up, implements the change in an isolated run, and opens a pull request linked back to the issue. Model labels like llm-claude-opus55 choose the agent, a base-<branch> label targets a non-default branch, and state labels track progress automatically. Several model labels on one issue produce separate branches and PRs to compare.
Planning splits a broad prompt into review-sized tasks, and Epic mode runs them as an ordered chain: each task becomes its own PR, and the next one starts only after the previous PR merges — optionally auto-merging on green CI. A large change ships as a sequence of reviewable diffs. Each branch stays focused.
Yes. The branch, commits, task page, and pull request give every agent the same visible starting point. A PR can start with one agent, get reviewed by another, and receive a follow-up from a third while the complete sequence remains in GitHub.
The planning flow builds a context preview before implementation starts. ProPR combines the files you name, files it believes are relevant, repository summaries, screenshots, attachments, and optional reference repositories. ProPR shows the model the code it needs without dumping the whole repository.
Plans, prompts, task history, commits, file changes, logs, model choices, cost, and follow-up activity stay attached to the task. Branches, commits, and PR conversations also remain in GitHub, where your existing tools can clone, grep, and diff them.
Yes. Agents run on your server, so a device only steers. The web UI is mobile-optimized for plans, tasks, logs, and dashboards, and because iteration happens in GitHub comments, the GitHub mobile app works as a remote control: label an issue, review the diff, leave a plain comment like Please handle 429 retries for temporary provider rate limits.
, and the follow-up commit lands on the same PR. See mobile and remote control.
Yes. The CLI covers plans, tasks, repositories, agents, settings, todos, logs, active tasks, and system status, so scripts and CI jobs go through the same review path.
You need a Docker-capable Linux host, GitHub access for the repositories you want to work on, and provider access for the agents you want to use. Plans, tasks, and task logs stay in your ProPR stack. The LLM calls go to the providers you configure. The quickstart takes you from install to the first agent PR.
The source, images, documentation, and live demo are public. Install with two commands and review your first agent PR.
When to reach for a live IDE agent, when to delegate to ProPR, and exactly how the architectures differ.