Prove the30% savingson your own machine.
Run the read-only scan against your local agent history. Atelier cuts tool calls by up to 90% and input/output tokens by up to 80%, and shows where paid context is leaking: broad search, whole-file reads, verbose tool output, repeated rediscovery, and missing proof.
No install required.
Tokens, calls, and loops.
Keep using Claude Code.
Local install. No account required to start. Keep using Claude Code normally.
$ curl -fsSL https://install.atelier.ws | bash- Read smarter: exact code ranges instead of file dumps
- Think sharper: less context spent rediscovering the same facts
- Talk less: compact output without losing commands or errors
- Never forget: session memory carries useful context forward
Why it saves
Talk less saves output tokens.
A tighter loop saves the run.
This is not a caveman prompt telling the model to grunt. Atelier changes the runtime around the model: what it can find, what it reads, what tools return, and what carries into the next turn.
BaselineThe agent asks for context and gets broad text back.
With AtelierRanked code search returns the likely files, symbols, and exact ranges.
BaselineWhole files and repeated reads fill paid context.
With AtelierRange reads, outlines, and summaries keep only the useful source in view.
BaselineTool output and assistant replies sprawl across the transcript.
With AtelierCompact results preserve commands, paths, errors, and diffs without filler.
BaselineEvery session rediscovers the same project facts.
With AtelierRecall and carry-forward context stop the loop from paying twice.
“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.”
While you work
See live savings add up in Claude Code
Screen recording of a Claude Code session using Atelier: the statusline shows cost, context usage, and token savings updating live as the agent works.
Proof, not a fake counter
The 30% claim comes from raw runs, not live badges.
Live badges show usage adding up. The benchmark claim comes from both arms running the same model, tasks, and environment. 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 | $5.34 | 40.2% cheaper |
* Atelier: 1 rep/task; baseline: public tbench.ai leaderboard, 5-rep average. † A few tasks timed out before cost capture — real spend, not zero. Full methodology → BENCHMARKS.md. Per-suite breakdown → Atelier vs vanilla Claude Code.
Retrieval quality
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.
Under the hood: tree-sitter parsing, a zoekt trigram index for the lexical channel, BGE-Code-v1 embeddings for the semantic one — full architecture →
| Tool | MRR | p95 | p100 |
|---|---|---|---|
★ Atelier +semantic (BGE) | 0.727 | 390ms | 1057ms |
★ Atelier lexical (default) | 0.676 | 134ms | 319ms |
serena | 0.401 | 3834ms | 269001ms |
ripgrep | 0.376 | 66ms | 522ms |
code-index-mcp | 0.343 | 377ms | 3830ms |
ast-grep | 0.312 | 1255ms | 8806ms |
universal-ctags | 0.237 | 1ms | 12ms |
At linux-kernel scale
4.5M lines, 42M tokens, one repo — still fast, and still finds the right answer.
| Mode | Indexing time |
|---|---|
| ★ Atelier lexical (default) | 2m 59s |
| ★ Atelier +semantic (BGE) | 23m 22s |
See what each one claims about itself, per tool →. Full 13-tool table, latency, and methodology → BENCHMARKS.md.
60 seconds · read-only
Do not take our 30% claim on faith.
Run the read-only scan against your own Claude/Codex history before you install. It prints the savings estimate from your local files.
$ curl -fsSL https://savings.atelier.ws | bashScans local agent sessions · temporary store · no login · no API keys.
Honestly
The trust is the audit trail.
Anyone can fake a counter. Atelier earns trust three ways: raw benchmark runs are committed, the savings scanner checks your own machine, and rows where Atelier does not win stay in the table. Terminal-Bench 2.1 is flat on accuracy (-0.2pp) and only cheaper on cost.
Read the raw runs on GitHub →Verify first. Install second.
Start the next session with proof, not promises.
SWE-bench Verified: 29.5% cheaper and +12.0pp more tasks solved. Same model, same tasks, same environment. Raw runs published.
Need it for a team? Large codebases, shared team memory, a cheaper agentic harness at scale.
Contact us about Enterprise