OpenRig explained: coordinate coding agents without losing the operator
OpenRig performance and cost: count seats, context and operator time
Measure coordination overhead without inventing model-speed numbers
What you will learn
- Separate startup from useful work
- Count model and review costs
- Measure retained work
Before you start
- Node 22 or 24 and tmux on a supported host
- One working provider account and a disposable repository
Turn the first owner/checker run into evidence for a safe rollout decision
Key takeaways
- Seat count is not throughput.
- Provider usage remains an external cost.
- Restoration outcomes need measurement.
Separate startup from useful work
Record daemon startup, tmux seat readiness, provider authentication and first useful response. Cold starts and restored sessions answer different questions; put them in separate rows.
The project supports multiple provider runtimes and a local kernel. No latency benchmark was run for this series, so a generic claim that two agents are twice as fast would be unsupported.
Count model and review costs
Each active provider session can consume its own model quota or subscription allowance. The checker also costs time and tokens, while a failed handoff can require the owner to reconstruct context.
Track completed, accepted changes and rework rather than message count. A queue entry and an exact-candidate review make the accounting more meaningful than the number of visible seats.
Measure retained work
Snapshots and stable seats may reduce recovery effort, but they do not promise perfect context continuity. Record which nodes resume, which start fresh and how much manual reconstruction is needed.
Include storage, host uptime, operator attention and provider limits. The local orchestration software does not remove third-party model costs.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Time cold launch and restored launch separately.
- 2
Record provider usage, checker time and accepted outcomes.
- 3
Track per-seat recovery after an interruption.
Copy-ready example
rig,provider,seats,cold_ready_s,restored_ready_s,accepted_changes,review_minutes,model_usage
,,,,,,,not-runFrequently asked questions
Does self-hosting eliminate model charges?
No. Selected providers can still charge or enforce usage limits.
Is a multi-agent speedup established here?
No. This series reports no runtime benchmark.
Sources
- OpenRig / README.mdSource checked 2026-10-04
- OpenRig / docs/reference/telemetry.mdSource checked 2026-10-04
- OpenRig / docs/reference/instance-layout.mdSource checked 2026-10-04
- OpenRig / docs/reference/getting-started.mdSource checked 2026-10-04