---
title: "OpenRig: Running Claude Code and Codex as One Agent Team"
date: "2026-10-03"
video_id: "7692250211711782174"
tiktok_url: "https://www.tiktok.com/@reporadarai/video/7692250211711782174"
creator: "@reporadarai"
category: "AI"
tags: ["openrig", "claude code", "codex", "multi-agent", "tmux"]
excerpt: "OpenRig defines a team of Claude Code and Codex agents in YAML and starts it with one command. What the starter does and what install changes on your machine."
---
OpenRig is an open-source command-line tool that defines a team of coding agents in YAML and launches them in your repository with one command, with Claude Code and Codex seats running side by side under tmux. A TikTok from @reporadarai says it "launches a two-agent team in your repository to drive one reviewed change," and that claim matches the project's docs. The video's star and download numbers went stale within days, and it scrolls past the section that matters most: the trust settings and hooks the install writes to your machine.

## What is OpenRig?

The project is [github.com/mvschwarz/openrig](https://github.com/mvschwarz/openrig), a TypeScript project under the Apache-2.0 license by Mike Schwarz, created on 2026-04-01. Its README tagline: "A harness wraps a model. A rig wraps your harnesses. Define your agent team in YAML, boot it with one command. Claude Code and Codex in the same rig, managed as one system."

The YAML format is called RigSpec. It declares pods, members, edges and continuity policies. A separate AgentSpec holds reusable agent blueprints, and a CULTURE.md file sets coordination norms. Seats work as addresses: `dev-owner@first-project` stays the same while the conversation in it changes.

## How do you install OpenRig?

The README gives:

```
npm install -g @openrig/cli
rig setup --dry-run
```

You need Node.js 22 or 24, tmux, and macOS or Linux. Native Windows is not supported and WSL2 is untested. Since version 0.6.0 the install check refuses Node 20. The npm package [@openrig/cli](https://www.npmjs.com/package/@openrig/cli) was first published on 2026-04-06 and has 54 versions.

To preview and start a starter team, the README runs `rig specs preview`, `rig up "$starter" --cwd . --plan` and `rig tui --shared`.

## What does the two-agent starter do?

The [getting-started guide](https://github.com/mvschwarz/openrig/blob/main/docs/reference/getting-started.md) says the two seats are an outcome owner and an independent checker. There are three starter rigs:

- **first-project** runs both seats on Codex, pinned to `gpt-6-astra`.
- **first-project-claude** runs both seats on Claude Code at its default model.
- **first-project-mixed** pairs a Claude owner with a Codex checker.

The models in the README's demo recording are real. `fable-5-1` is Anthropic's [Claude Fable 5.1](https://www.anthropic.com/claude/fable), announced on 2026-09-01. `gpt-6-astra` is an OpenAI model listed on [OpenRouter](https://openrouter.ai/openai/gpt-6-astra).

The README also has an Agent-Managed Software section about packaging software beside the agents that run it. Its example, `secrets-manager`, is a HashiCorp Vault instance run by a specialist agent, and it requires Docker. I did not install or run OpenRig, so I cannot say how well the starter performs on a real repository.

## Can OpenRig recover agents after a reboot?

Yes. `rig down --snapshot` saves the topology, and `rig up <name>` restores from the latest snapshot. Restore reports each seat as resumed, fresh or failed. After a reboot, the guide says to open `rig`, start the daemon if needed, and select the existing rig. Resume is the default.

Two caveats apply. The README's own demo, playing under the video's "organized and recoverable" caption, shows one seat failed, four in an unknown state, and a health line reading "canonical health records could not be read." The README also says sessions you adopt while they are already running may need a restart to pick up new runtime config.

## What does OpenRig change on your machine?

The README has a table titled "What OpenRig changes on your machine," and the video scrolls past it without comment. The changes include:

- An OpenRig block written into `~/.tmux.conf`.
- On macOS, an optional cmux install that switches `automation.socketControlMode` from `cmuxOnly` to `automation`.
- A skill seeded into `~/.claude/skills` and `~/.agents/skills`, plus Codex hook configuration and trust records, all written when the daemon starts.
- Managed launches that pre-trust the workspace, run Claude with `--permission-mode acceptEdits`, and run Codex with `-s workspace-write`.

The README says these include "trust settings and executable hooks" and tells you to back up the affected files first. Full-bypass modes are off by default. Activity relays send event type, seat identity, timestamps and session identity to the local daemon, and exclude prompt text and tool arguments.

## How popular is OpenRig?

The video shows 2,437 stars, 2.5k weekly downloads and version 0.6.1. On 2026-10-03 the repo had 4,645 stars and 321 forks, weekly downloads read 4.6k, and v0.6.4 had shipped on 2026-10-02. npm dates v0.6.1 to 2026-09-29, so the screen capture came from the last days of September.

## OpenRig vs claude-squad

Running several terminal coding agents under tmux is not new. [claude-squad](https://github.com/smtg-ai/claude-squad), which manages Claude Code, Codex, OpenCode and Amp sessions, was created on 2025-03-09 and has 8,565 stars. OpenRig adds a declarative YAML topology, stable seat addresses, snapshot and restore, `rig discover` and `rig adopt` for bringing existing tmux sessions under management, and Claude Code and Codex seats working as peers in one rig.

## Bottom line

OpenRig does what the video says: it starts an owner agent and a checker agent in your repo from a YAML definition, with Claude Code, Codex or both. Read the "What OpenRig changes on your machine" table and back up your tmux, Claude and Codex config before installing.

## Sources

- [mvschwarz/openrig on GitHub](https://github.com/mvschwarz/openrig) - repo, license and star count.
- [OpenRig README](https://github.com/mvschwarz/openrig/blob/main/README.md) - RigSpec, snapshots and the machine changes table.
- [OpenRig getting-started guide](https://github.com/mvschwarz/openrig/blob/main/docs/reference/getting-started.md) - owner and checker seats and starter rigs.
- [@openrig/cli on npm](https://www.npmjs.com/package/@openrig/cli) - versions and Node requirement.
- [OpenRig vs Claude managed agents](https://openrig.dev/compare/claude-managed-agents) - the local daemon, SQLite and tmux architecture.
- [Claude Fable 5.1](https://www.anthropic.com/claude/fable) - the model in the demo.
- [GPT-6 Astra on OpenRouter](https://openrouter.ai/openai/gpt-6-astra) - the OpenAI model slug.
- [smtg-ai/claude-squad](https://github.com/smtg-ai/claude-squad) - earlier tmux-based agent manager.


## Video description and verification notes

**A 26-second vertical clip at 720x1280, posted 2026-10-03 by @reporadarai (channel "Repo Radar AI"), carrying 394 views, 15 likes, 0 comments, 3 reposts and 18 saves at capture, over the audio track "original sound - Repo Radar AI". I analyzed 13 extracted frames and a 67-word transcript. Twelve of the thirteen frames are a slow autoscroll through the rendered README of a single GitHub project, OpenRig, with burned-in captions over it. The README header shows four shields badges: npm v0.6.1, downloads 2.5k/week, license Apache-2.0, and GitHub Stars 2.4k. Visible README text includes the tagline "A harness wraps a model. A rig wraps your harnesses. Define your agent team in YAML, boot it with one command. Claude Code and Codex in the same rig, managed as one system," the install block `npm install -g @openrig/cli` followed by `rig setup --dry-run`, a requirements line naming Node.js 22 or 24 plus tmux on macOS or Linux, a three-row table of starter teams (Two Codex agents / Two Claude agents / Claude owner + Codex checker, mapped to `first-project`, `first-project-claude`, `first-project-mixed`), a command block running `rig specs preview`, `rig up "$starter" --cwd . --plan` and `rig tui --shared`, and a long "What OpenRig changes on your machine" table covering npm installation, `rig setup`, daemon startup, rig/seat launch and explicit permission configuration. Embedded in the README is a terminal recording of the TUI topology table showing a rig named `v-openrig-build` with pods `dev`, `orch`, `review`, `dev50` and `review50`, seats named driver, guard, qa, planner, advisor, lead, successor and rebuild-historian, a MODEL column alternating `fable-5-1` and `gpt-6-astra`, context percentages from 20% to 88%, states reading idle, working, unknown and failed, a HEALTH line reading "Unknown · Unavailable · canonical health records could not be read," and a status bar reading "17 seats · 6 working · 1 need attention · 0 open rows." The final frame is a dark end card: "FIND IT AT github.com/mvschwarz/openrig", a star pill reading 2,437, a truncated install line, and the question "Would you run your next coding project as a persistent agent team?"**

## The repo slug, the author and the license all check out

The end card's slug is correct. [github.com/mvschwarz/openrig](https://github.com/mvschwarz/openrig) resolves to a TypeScript project licensed Apache-2.0, owned by Mike Schwarz (GitHub `mvschwarz`, X handle `@_feralmachine`, 305 followers), created 2026-04-01 and last pushed 2026-10-03. The GitHub topics on the repo are the same tags the TikTok description uses: `agent-harness`, `agent-orchestration`, `agent-skills`, `ai-coding`, `claude-code`. The npm package is [`@openrig/cli`](https://www.npmjs.com/package/@openrig/cli), first published 2026-04-06, 54 versions to date, engines `node: >=22`.

The two model names in the embedded TUI recording are real, not placeholder strings. `fable-5-1` is Anthropic's [Claude Fable 5.1](https://www.anthropic.com/claude/fable), API identifier `claude-fable-5-1`, announced 2026-09-01. `gpt-6-astra` is OpenAI's model, listed on [OpenRouter](https://openrouter.ai/openai/gpt-6-astra) as `openai/gpt-6-astra`.

## Every number on screen was stale within days of posting

The badges and the end card are a snapshot, and the clip presents them without a date. At the time I wrote this, 2026-10-03, the live values are: 4,645 stars against the 2,437 on the end card, 4.6k weekly npm downloads against the 2.5k badge, and version 0.6.4 against the v0.6.1 badge. The npm registry dates v0.6.1 to 2026-09-29 and v0.6.4 to 2026-10-02, which puts the screen capture in a four-day window at the end of September. The repo roughly doubled its star count inside that window. Treat the 2.4k figure as a floor, not a current reading.

## The two-agent claim is accurate and more specific than the voiceover says

The voiceover says OpenRig "launches a two-agent team in your repository to drive one reviewed change." The README's own sentence is "install, launch a two-agent team in your repository, and get one reviewed change," so that is a direct lift. The [getting-started guide](https://github.com/mvschwarz/openrig/blob/main/docs/reference/getting-started.md) is more precise about what the two seats are: an outcome owner and an independent checker. `first-project` runs both seats on Codex pinned to `gpt-6-astra`. `first-project-claude` runs both on Claude Code at the configured native default. `first-project-mixed` is a Claude owner with a Codex checker. The clip's on-screen table shows this; the voiceover does not.

The YAML claim is also correct. OpenRig calls the format RigSpec, a declarative definition of pods, members, edges and continuity policies, with a separate AgentSpec for reusable agent blueprints and a CULTURE.md for coordination norms. Seats are addresses: `dev-owner@first-project` stays stable while the conversation occupying it changes.

## "Recoverable after a reboot" is documented, and the demo footage shows its limits

Recovery is a real feature, not marketing. `rig down --snapshot` captures topology state, `rig up <name>` restores from the latest snapshot, and restore reports per-node outcomes as resumed, fresh or failed. The getting-started guide's reboot path says to open `rig`, start the daemon if needed, and select the existing rig and seats, with resume as the default and a fresh conversation as a separate decision. Two things qualify the clip's "organized and recoverable" caption. First, the README's own embedded recording, playing under that exact caption, shows a seat named `rebuild-hist…` in a `failed` state, four seats in `unknown`, and a health line reading "canonical health records could not be read." Second, the README says adopted sessions already running may need a restart before they pick up newly written runtime config.

## The video omits what installing this writes to your machine

The clip scrolls past the "What OpenRig changes on your machine" table without ever reading it aloud, and it is the most consequential section of the README. OpenRig writes an OpenRig block into `~/.tmux.conf`; on macOS it can install cmux and flip `automation.socketControlMode` from `cmuxOnly` to `automation`; daemon startup seeds a skill into `~/.claude/skills` and `~/.agents/skills` and writes Codex hook configuration plus trust records before any rig launches; managed launches pre-trust the workspace, run Claude with `--permission-mode acceptEdits`, and run Codex with `-s workspace-write`. The README states plainly that these include "trust settings and executable hooks" and tells you to back up the relevant files first. Full-bypass modes are off by default. Activity relays post event type, seat identity, timestamps and native session identity to the local daemon, excluding prompt text and tool arguments.

## The category is older than this repo

The clip opens by framing scattered terminal sessions as the problem OpenRig solves, which reads as a first-mover claim. Managing several terminal coding agents under tmux predates OpenRig by more than a year. [smtg-ai/claude-squad](https://github.com/smtg-ai/claude-squad), which manages Claude Code, Codex, OpenCode and Amp sessions, was created 2025-03-09 and now has 8,565 stars, against OpenRig's creation date of 2026-04-01. What OpenRig adds on top of that category is the declarative YAML topology, stable seat addresses, snapshot and restore, `rig discover` and `rig adopt` for pulling existing tmux sessions under management, and a documented path for Claude Code and Codex seats acting as peers in one rig.

## Key Takeaways

- The repo slug on the end card is correct: `github.com/mvschwarz/openrig`, Apache-2.0, TypeScript, created 2026-04-01, author Mike Schwarz.
- Correction: the 2,437 stars on the end card read 4,645 on 2026-10-03, and the 2.5k/week downloads badge now reads 4.6k/week.
- Correction: the v0.6.1 badge in the clip is two patch releases behind; v0.6.4 shipped 2026-10-02, which dates the screen capture to roughly 2026-09-29 through 2026-10-01.
- The two agents are an outcome owner and an independent checker, and the three starter rigs are `first-project` (two Codex seats pinned to `gpt-6-astra`), `first-project-claude` (two Claude Code seats), and `first-project-mixed`.
- "Package software beside the agents managing it" is verbatim from the README's Agent-Managed Software section. The shipped example is `secrets-manager`, a HashiCorp Vault instance operated by a specialist agent, and it requires Docker.
- Requirements the clip shows but does not say: Node.js 22 or 24, tmux, macOS or Linux. Native Windows is unsupported and WSL2 is untested. Node 20 is refused by the install check since 0.6.0.
- Install writes trust settings and executable hooks, including a `~/.tmux.conf` block, Codex hook config, and `acceptEdits` as Claude's permission mode in managed launches.
- Prior art: `smtg-ai/claude-squad` has done tmux-based multi-agent session management since 2025-03-09, 13 months before this repo existed.
- Both models in the demo footage are real: `claude-fable-5-1` (announced 2026-09-01) and `openai/gpt-6-astra`.
- Unverified: the exact date the star count crossed 2,437, because GitHub caps stargazer pagination well below that depth without authentication. Unverified: the GPT-6 Astra announcement date, since openai.com and the CNBC report both returned 403 to a direct fetch. Unverified: whether the two-agent starter actually produces a reviewed change on a real repository, since I did not install or run it.

## Resources

- [mvschwarz/openrig on GitHub](https://github.com/mvschwarz/openrig) — confirms the slug, Apache-2.0 license, TypeScript, 2026-04-01 creation, 4,645 stars and 321 forks on 2026-10-03.
- [OpenRig README](https://github.com/mvschwarz/openrig/blob/main/README.md) — source of the YAML/RigSpec description, snapshot and restore behavior, the Agent-Managed Software section, and the full "what OpenRig changes on your machine" table.
- [OpenRig getting-started guide](https://github.com/mvschwarz/openrig/blob/main/docs/reference/getting-started.md) — confirms the owner-and-checker seat pair, the three starter rigs with their pinned models, and the post-reboot resume path.
- [@openrig/cli on npm](https://www.npmjs.com/package/@openrig/cli) — confirms 54 published versions, first publish 2026-04-06, v0.6.4 on 2026-10-02, `node: >=22`.
- [OpenRig vs Claude managed agents](https://openrig.dev/compare/claude-managed-agents) — confirms the self-hosted local daemon, SQLite and tmux architecture and the cross-runtime peer claim.
- [Claude Fable 5.1](https://www.anthropic.com/claude/fable) — confirms `claude-fable-5-1` as a real API identifier announced 2026-09-01, matching the `fable-5-1` strings in the demo TUI.
- [openai/gpt-6-astra on OpenRouter](https://openrouter.ai/openai/gpt-6-astra) — confirms `gpt-6-astra` is a real OpenAI model slug.
- [smtg-ai/claude-squad](https://github.com/smtg-ai/claude-squad) — confirms tmux-based multi-agent terminal management existed from 2025-03-09, predating OpenRig.
