OpenRig Runs Claude Code and Codex as One Agent Team
A.I. / news
OpenRig Runs Claude Code and Codex as One Agent Team
The free, self-hosted tool picked up 114 stars in a single day while Anthropic charges 8 cents an hour for its own hosted version.
An open-source tool called OpenRig gained 114 stars in a single day on Sept. 28, landing on GitHub's trending page for a project that runs Anthropic's Claude Code and OpenAI's Codex as one coordinated team of coding agents.
The repository, built by a developer using the handle mvschwarz, has drawn 1,300 stars and 117 forks under an Apache 2.0 license. OpenRig lets a developer define a group of agents in a YAML file, boot the whole set with a single rig up command, and manage them through a terminal interface that shows which agents are running, how they are connected, and what each one is doing. The system runs as a local daemon on top of tmux, backed by a SQLite database, and exposes an MCP server so agents can inspect and change their own topology.
What running it costs instead
OpenRig's own comparison points at Anthropic's hosted alternative, Claude Managed Agents, which Anthropic's pricing page lists at $0.08 per session-hour of active runtime on top of standard API token rates. OpenRig charges nothing beyond the cost of the machine it runs on, but it only coordinates agents a developer already has installed and pays for separately through their own API keys.

| OpenRig | Claude Managed Agents | |
|---|---|---|
| Hosting | Self-hosted, local machine | Anthropic-hosted, cloud |
| Cost | Free, Apache 2.0 license | $0.08 per session-hour of runtime |
| Engines | Claude Code and Codex together | Claude only |
The project requires Node.js 20, 22 or 24, and tmux is mandatory rather than optional. Native Windows is not supported, and the documentation says WSL2 has not been tested. Installing it writes trust settings and executable hooks into configuration files including ~/.tmux.conf, ~/.claude.json and ~/.codex/config.toml, and the author tells users to review and back up those files first.
What early users are running into
A Show HN discussion of the project drew a critique of multi-agent swarms generally: one commenter wrote that without supervision, "even the most advanced models degrade quickly," and that a single early mistake can snowball as later agents build on a wrong approach. The same commenter said agent swarms were most useful for mapping a project at the start rather than running it unsupervised to the end.
The project's author responded that Claude Code and Codex are meant to complement each other in the harness, not duplicate one another: Claude Code for understanding intent, Codex for more conservative execution that produces fewer mistakes. The author recommended test-driven development as a gate on agent output and pointed to topology persistence, saving a running set of agents and restoring it later by name, as the feature a raw tmux setup does not provide.
The documentation describes OpenRig as still early, with what it calls "the core primitives" in place and the setup described in the README functioning, but with gaps the project has not yet closed. It ships no built-in way to track combined API spending across the agents it coordinates, so a developer running a large rig still has to add up token costs from Anthropic and OpenAI separately. Pairing two full harnesses does not remove either one's own overhead: a UC Berkeley study found that Claude Code's scaffolding alone can cost roughly twice as many tokens as a minimal harness for a similar success rate. Other vendors are competing on that overhead directly; AWS's own Strands Harness says it beat Claude Code's per-trial cost by more than four times in its own Terminal-Bench 2.1 testing by trimming the scaffolding rather than adding a second engine.
Sources
- 03Pricing
More in A.I.
- 01Runta's 27,280-Run Harness Eval Finds the Model Explains 92.8% of Agent CostFrontierHarness v1.1 ran 22 coding-agent harnesses against five open models, and Codex CLI went from first place to sixteenth when the model changed.
- 02Qwen3.8-27B Has 6.9 Million Downloads on Hugging Face. Artificial Analysis Rates It Slow and VerboseAlibaba's Apache 2.0 dense model beats its size class on an independent index of 20, and takes 38.8 minutes a task against a 12.6-minute median.
- 03Claude Code Team Publishes Its Verification Rules: 39 Checks, a Three-Round Limit and Observe ModeA Nov. 16 post says about 90% of the team's code is written and reviewed by Claude, and lays out the gates that keep it shippable.
- 04Inkling 2 Small Weights Land on Hugging Face Under Apache 2.0, and the Card Asks for 500GB of VRAMThinking Machines opened the weights a week after launching on Tinker. The safety section still says release is pending.