We Open Sourced R-CLI, the Coding Harness Above Every Published Terminal Bench 2.1 Result

R-CLI is now open source. MIT license. The whole harness. Repo: https://github.com/Backboard-io/Backboard-R-CLI First, the ask We are giving away the harness behind the highest scores ever recorded on Terminal Bench 2.1. Free, MIT, running on your own keys in about two minutes. The receipts are below, with public verifier logs for every claim. The entire price: Star the repo. Stars are how open source projects get found. Thirty seconds. Share it with one developer who pays too much for AI coding. That is it. This is a win for the community and for open source, and we want it loud. Three ways to run it One command on macOS or Linux: curl -fsSL https://app.backboard.io/api/cli | sh Windows PowerShell: irm https://app.backboard.io/api/cli/windows | iex Or build from source, because now you can: clone the repo, bun install --frozen-lockfile, then bun run build for a native binary on your OS. Bun, Git, and ripgrep are the only prerequisites. Here is why this matters, and why we gave it away. The thesis: the harness is the dominant variable The industry story is that coding performance lives in the model. Pick the biggest frontier model, pay the bill, get the results. Our runs say otherwise. Take the same model, change only the harness, and the score moves more than a model generation's worth. R-CLI is built on the RLM, a recursive coding engine. The parent agent delegates work off-model to child agents that run in bounded contexts. Each child gets a sharp, small slice of the problem instead of one giant context window accumulating noise. Work comes back up, context gets compressed, and the loop continues. That structure is why the same model scores higher inside R-CLI, and why it does it with 10 to 30 percent fewer tokens. The numbers, with receipts Two published runs from our technical reports. Per-task verifier logs are public at https://github.com/Backboard-io/Backboard-R-CLI-Terminal-Bench-2.1-Results Same model, better harness. Claude Opus 4.8, full Terminal Bench 2.1 submission shape: 89 tasks, 5 attempts each, 445 trials, no concessions. R-CLI scored 84.04 percent (374 of 445). Claude Code's published score with the same model is 78.9 percent. That is +5.14 points with exactly one variable changed. The result sits above all 17 published entries. It has not yet been submitted for judge review, so treat it as a published run with public logs, not a leaderboard entry. No frontier model in the run. Moonshot AI's Kimi K3 via OpenRouter, with DeepSeek V4 Flash handling subagent work. Single attempt per task: 91.0 percent (81 of 89). Total cost for the entire run: $28.72. About 32 cents per task. Caveat carried in the report itself: single-attempt runs have a roughly ±3 percent confidence interval and this is not a judged submission. Even at the bottom of that interval, an open-weight model in this harness is playing at the top of the benchmark for a rounding error of the usual cost. One run shows the harness lifts a frontier model past its own tooling. The other shows the harness makes frontier-level coding nearly free. Same engine both times. What you actually get Terminal-native coding agent: inspects files, edits, runs commands, searches the web The recursive engine: off-model delegation, bounded child contexts, context compression Hundreds of pre-loaded skills, browsable with /skills, plus your own in .agents/skills/ Hundreds of MCP connections, managed with /mcp Permission modes: manual, acceptEdits, bypass. You decide what it can touch Checkpoints: undo and redo agent changes, reviewable diffs Session logging, one-shot mode, JSON output for scripting, hooks via /hooks ## Any model. Your keys. R-CLI is model-agnostic, with a path to hundreds of models. Bring your own key for OpenAI, OpenRouter, Anthropic, or Grok, or sign in with Backboard. Switch with /model. When a better or cheaper model ships next month, you change one setting. No re-platforming, no new tool. The part that matters most: run it where your code lives Every closed coding tool ships your source to a vendor API. That is the deal, whether the marketing says so or not. Pair R-CLI with an open model you host yourself and that deal is off. Frontier-level coding where your code never leaves your infrastructure. The Kimi K3 run is the proof that this path is not a performance downgrade. And now the harness itself is MIT. You can read it, audit it, fork it, and run it inside your own walls. Nobody can policy-block it, deprecate it, or reprice it out from under you. Why open source it Because the thesis demands it. If we claim the harness is the dominant variable, the harness should be inspectable. Benchmarks with public verifier logs were half of that. Open code is the other half. And because ownership is the whole point of what we build. Own the model. Own the harness. Run both in your own cloud. The ask, one more time We gave away the harness. You star and share it. That is the trade. https://github.com/Backboard-io/Backboard-R-CLI Then install it, break it, open issues, send PRs: curl -fsSL https://app.backboard.io/api/cli | sh Docs: https://docs.backboard.io