llms.txt
Compare with other sites' llms.txt files →# Relay
> Relay is the meta harness for coding, with better results than Claude Code on the same model. Relay races Claude Code and Codex on your hard tasks and hands back only the change your own tests prove, on the plans you already have. On Terminal-Bench 2.1, with Claude Haiku 5.5 for every entry, the Relay meta-harness solved 86.2% of 87 tasks and Claude Code 2.1.300 alone 83.3% (difference +2.9 points, 95% interval −1.1 to 6.9), from round 1’s saved runs.
Relay races Claude Code and Codex on your hard tasks and hands back only the change your own tests prove, on the plans you already have. macOS, Linux, WSL. Free. Free account, one click, right after install.
One command installs it:
curl -fsSL https://relayevals.com/install.sh | bash
That is the only thing a person installs. Relay is a harness you run yourself; there is no hosted execution and no resold inference.
## Results
- On Terminal-Bench 2.1, with Claude Haiku 5.5 for every entry, the Relay meta-harness solved 86.2% of 87 tasks and Claude Code 2.1.300 alone 83.3% (difference +2.9 points, 95% interval −1.1 to 6.9), from round 1’s saved runs.
- Same tasks and model for every entry. The meta-harness kept Claude Code 2.1.300’s or Relay 0.6.118’s saved attempt on each task, by its picker; Relay 0.6.118 alone solved 71.8%. Round 1, 16 to 17 Oct 2026.
- Source: https://github.com/Relay-harness/relay/tree/main/benchmarks/scoreboard
## Today
- One command installs the harness on macOS, Linux and WSL. A free Relay account, made in one click right after install, with GitHub, Google, Agent ID or email.
- /login connects a Claude, ChatGPT or GitHub Copilot subscription, or an API key.
- Sessions run in a resident worker and survive a closed terminal; relay attach reconnects.
- /goal and /autonomous, with completion gates and limits on turns, tokens, time and spend.
- Recursive subagents from the Python kernel.
- Memory and skills through /refine; relay learning lists them.
- Scheduled and recurring prompts with relay schedule.
- MCP servers through Python skills.
- relay queue: a backlog to verified branches and pull requests, one worktree per task; one repository on Free, any number on Pro.
- Relay re-runs test, lint and build itself, in a clean checkout or a sandbox, before a pull request exists.
- A report per task, committed on its branch as the pull request body, and a report per run.
- The model router: a top model plans and reviews, lighter models code, chosen per task and recorded (Pro).
- Checks in fresh Relay sandboxes (Pro, 300 minutes a month).
- Brakes: stop all, pause or quarantine a repository, in relay panel and the terminal.
- The dashboard at relayevals.com/control: connect apps (MCP servers), AI models and secrets to your computer, and see that it is in sync, on every plan.
- The learning ablation: relay policy --learning-ablation on splits tasks between learning off, on and frozen.
- A signed-in machine sends numbers: each session's cost and tokens, and, with remote sync on, task titles and statuses. Never a brief, a log, a diff or a prompt.
- 1,244 models from 31 providers in the catalogue bundled with Relay 0.6.122, and new ones arrive between releases: relay reads the model catalogue from relayevals.com at most once an hour, and a catalogue can never change where a key is sent.
- Relay is closed source. The open-source components it includes keep their own licences and copyright notices, which ship with it.
## Prices
- Free: $0 forever, no card required. The whole agent, on your own machine. No card required. A free Relay account, made in one click right after install, with GitHub, Google, Agent ID or email.
- Pro: $12 per month, billed yearly at $144. Per month billed yearly ($144), or $15 month-to-month. The task queue across any number of repositories, unattended after you close the terminal; a top model plans and reviews while lighter models code, chosen per task; and each task’s checks run in a fresh Relay sandbox, 300 minutes a month included.
- Team: $20 per developer per month, billed yearly. Per developer per month billed yearly ($240 per seat), or $25 month-to-month, from 2 seats. Everything in Pro for every member, 600 sandbox minutes per seat pooled, one organization baseline your admin publishes for the team’s CI checks, which machines are signed in and what each one spends, and 365 days of history, downloadable as CSV.
- Enterprise: From $20k a year, by contract. Everything in Team, with the team audit log and its CSV export, 1,500 sandbox minutes per seat pooled, invoicing, help with your security review, and direct support. On contract: SAML/OIDC single sign-on, SCIM, retention settings, data location, a private sandbox broker, and an SLA. Write to arnav@relayevals.com.
Prices verified 2026-09-18. Founding prices: subscribe now and yours is locked while you stay subscribed. Your model. Your keys. Your machine. Tokens are never billed.
## Docs
- [The harness reference](https://relayevals.com/docs): every relay command, subcommand and flag, generated from the harness's own command registry.
- [Quickstart](https://relayevals.com/docs/start): install, start it in a repository, connect a model, type the task.
- [Pricing](https://relayevals.com/pricing)
- [Changelog](https://relayevals.com/changelog)
- [The dashboard](https://relayevals.com/control/apps): connect apps, AI models and secrets to your computer, and see that it is in sync.
- [llms-full.txt](https://relayevals.com/llms-full.txt): the harness's complete command surface, introspected from the shipped binary.