Offrun runs Claude Code and Codex side by side: worktrees, logins, limits

The project launched as a Show HN on 2026-10-03. The landing page at offrun.dev describes it as a layer around existing harnesses, and its FAQ says it does not replace Claude Code or Codex but runs them. Pi and OpenCode are listed as coming soon. For an engineer deciding whether to put this between themselves and their agents, the questions are what gets isolated, where credentials live, and what you lose by not driving each CLI directly.

The isolation unit is the git worktree. Every agent working in a project gets its own worktree, so two agents on the same repo do not touch the same files. The product example runs a refactor and a bug fix at once and leaves the merge to you. In the demo, a flaky-test fix reports that it changed only test/upload.test.js on its own branch and passed 50 runs in a row.

Workspace state is stored in the repository, not in a hosted service. Chats, memory and worktrees live under .offrun and .worktrees in the project folder, and the page says Offrun keeps no copy. Teams will need to decide whether those directories go into their ignore rules or get committed. The page invites you to version the memory files, but it says nothing about whether chat history is meant to be committed.

Execution isolation is not described. The page says the app runs the CLIs on your Mac, signed in as you. It does not mention containers, a VM, filesystem permissions beyond the worktree, or network egress controls. On what the sources describe, a worktree is a way to keep agents from editing the same files, not a security boundary.

Offrun requires no new account and no API key. You connect harness accounts under Settings, then Agents, and you can add a second account per harness. The demo shows Claude Code with a Default login at sam@studio.dev and a Work login at sam@acme.co. It also shows Codex signed in with ChatGPT, AGY signed in with Google, and Grok Build signed in with X. The page says each agent runs in its own folder and that Offrun never sees your password.

The data-flow claim is specific. Prompts and code go from the Mac to Anthropic, OpenAI, Google or xAI on your own login, and Offrun has no server in that path and never handles your keys. The sources do not explain how separate logins for the same CLI are kept apart on disk.

The pricing section also advertises free credits for 100+ connectors, which let agents find an email, file a GitHub issue or post to Slack. The no-server claim is made for the prompt path, and the connector path is not described. Check how connector traffic is routed before you enable connectors on a work machine.

Peer review uses one of your own agent accounts as a reviewer, and you choose the model in the model picker. Findings appear in the chat while the work is still uncommitted. If you ask for fixes, they land in your message box for you to read before anything is sent. The FAQ says peer review cannot change your code, and the page says nothing merges on its own. In the demo, Codex reviews this retry-backoff change:

The review is marked 'changes requested' and flags a warning at src/retry.js:2. The 30-second cap is longer than the gateway's 25-second timeout, so the last retry is cut off. The page says review runs on your own agent accounts, with no second subscription and no bot on your pull requests. It does not say how much usage a review consumes on the reviewing account.

Limit handling also runs across your accounts. When one login hits its usage limit, Offrun moves the chat to a login that still has room and tells you. You send your message again, and the new login picks up the whole conversation. When you are out of logins, it offers another agent. The demo has Codex continuing a chat after Claude Code hits its limit, with a note that Claude Code resets at 3am.

The sources do not say how a conversation built in one harness is replayed into another. They also do not say what tool state, if any, carries over.

The workspace keeps repo-wide conventions in one place, and each agent's goal, plan and dead ends in another. The api-gateway example has three files: conventions.md, which every agent reads; plan.md, for one feature's approach; and dead-ends.md, which records what was tried and failed. Agents read that context at the start of each session and write back to it as they work, so the fourth session knows what the first one already tried.

Because these are plain files in the project folder, you can open, edit and commit them like any other file. The same property means anything an agent writes back will be read by later sessions unless someone reviews it. Diffing these files is the obvious control.

The remaining features are about convenience. A preview pane shows changed files, diffs with line numbers, rendered markdown and tables, images, PDFs and a small browser with an address bar. Dictation transcribes on-device and needs a network connection only for setup. It keeps identifiers spelled the way your code spells them, with a vocabulary list that maps 'retry backoff helper' to retryBackoffHelper, and optional voice samples.

The top question was how Offrun differs from Herdr. A reply written for the product said Herdr lets you run agents anywhere, Offrun helps you manage them, and Herdr integration is coming. Another commenter pointed out that Herdr runs on Linux, Windows and Intel Macs, and said it has reportedly raised $6 million.

A product-side reply said Offrun was released for Mac on both Apple Silicon and Intel, with Windows and Linux to follow. That conflicts with the landing page, which says Apple Silicon only. The same reply said integrations for Herdr, Pi and OpenCode, plus a way to add a custom harness, are being worked on.

One commenter asked to see per-account rate-limit headroom before dispatching, not after getting stuck mid-task. The reply did not address pre-dispatch visibility. It said users can switch to another account in one click and resume.

Others compared Offrun with Paseo, Orca, Goose, T3 Code, Conductor, JetBrains Air and more than half a dozen others. The product-side answer was a difference in design philosophy: other meta-harnesses focus on agent-in-the-loop, and Offrun focuses on human-in-the-loop. The reply said Offrun would be feature-complete in the next two months, with a cloud product and a mobile app launching soon. One commenter argued that tools with tighter single-harness integration may work better because ACP is fairly bare. The sources do not say whether Offrun uses ACP.

The launch leaves several engineering questions unanswered. On each one below, the landing page and the thread are either silent or contradict each other:

Until these are documented, the safe reading is narrow. Offrun is a free local front end for four vendor CLIs that adds per-agent worktrees, a reviewer that cannot change your code, and file-based memory. Any security and execution guarantees still come from the harnesses underneath.