Atelier
Runtime.
Open-source runtime for coding agents. Give Claude Code, Codex, Cursor, and other coding agents a shared runtime for tools, context, skills, observability, cost control, and safer execution inside real codebases.
Cheaper on every suite we tested, and more correct on every multi-rep one — same model, same tasks, same environment. Every raw run published.
$ curl -fsSL https://install.atelier.ws | bashApache-2.0 · local-first · no account to start
Watch the savings add up
Statusline: atelier | model · ctx used % | $cost (I input / C cache / O output) ↓ saved ♻ context carry. Numbers are an illustrative session, not a benchmark.
Same model. Every suite. Receipts published.
Each benchmark runs both arms on the identical model, tasks, and environment — the only variable is the runtime. Every raw run is committed to the repo.
| Benchmark | Baseline | Atelier | Δ correct | Baseline $ | Atelier $ | Δ cost | Savings |
|---|---|---|---|---|---|---|---|
SWE-bench Verified 50 tasks × 5 reps | 80.8% | 92.8% | +12.0 pp | $234.84 | $165.45 | 29.5% cheaper | |
SWE-bench Lite 10 tasks × 3 reps | 93.3% | 100% | +6.7 pp | $12.38 | $10.79 | 12.9% cheaper | |
SWE-bench Pro 10 tasks × 5 reps | 88.0% | 90.0% | +2.0 pp | $39.01 | $30.61 | 21.5% cheaper | |
Exploration 7 large repos × 5 reps | — | — | — | $19.11 | $6.29 | 67% cheaper | |
Terminal-Bench 2.1 89 tasks vs public leaderboard* | 78.9% exp. | 78.7% | −0.2 pp | $96.76 | $69.52† | 28.1% cheaper† | |
Telegraphic Q&A 20 prompts × 5 reps | — | — | — | $8.93 | $6.81 | 23.7% cheaper |
* Model: Opus-4.8, Atelier: 1 rep/task; baseline: public tbench.ai leaderboard, 5-rep average. † A few tasks timed out before cost could be captured — real spend, not zero. Full methodology and every raw run → BENCHMARKS.md.
It finds the right code — better than grep.
Mean Reciprocal Rank across ~7,200 query/answer pairs on 14 repos — every tool scored on the same corpus and the same query kinds. Higher is better.
…and fast: Atelier's default lexical search answers in 134ms p95 — versus serena's 3.8s. Full 13-tool table, latency, and methodology → BENCHMARKS.md.
A leaner reply saves output tokens. A tighter loop saves the run.
Everyone’s teaching agents to talk less — that’s the easy 5%. Atelier compresses the other 95%: what the agent reads, calls, and re-reads. Code, commands, and errors stay byte-for-byte exact.
“I looked into the failing test and it seems like the flakiness is caused by the retry logic using a real clock. The test sleeps for 100ms and then asserts that exactly three retries happened, but under CI load the timing can drift, which makes the assertion fail intermittently. I’d recommend injecting a fake clock so the test becomes deterministic.”
“Root cause: retry test uses a real clock — 100ms sleep + exact 3-retry assert drifts under CI load. Fix: inject a fake clock; test becomes deterministic.”
Don’t take our benchmark’s word for it.
Point it at your own agent history and see what routable calls would have cost less.
$ curl -fsSL https://savings.atelier.ws | bashScans your local sessions · temporary store · no login, no API keys.
We publish the benchmark where we lose.
Terminal-Bench 2.1: Atelier is flat on accuracy (−0.2pp) and wins only on cost (28.1% cheaper). It’s in the table, not the footnotes. Every headline number links to committed raw runs. No cherry-picking — the credibility is the product.
Read the benchmark on GitHub →Run every coding agent on one runtime.
In our SWE-bench run, Atelier cut 2,626 turns and about $69 off the baseline. A star costs zero tokens. Fair trade.