Claude Code
Codex
opencodeOne shot to theright code.A faster agent.
Atelier gives Claude Code one-shot code search: the right files, symbols, and exact line ranges in a single call. Fewer round-trips means 37.7% fewer turns and 23.7% less wall-clock on SWE-bench Verified -- and about 30% lower cost.
30 seconds. Keep using Claude Code.
atelier savings: wasted calls and tokens in your own history.
Turns and cost fall live in the statusline.
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: no turns burned rediscovering the same facts
- Talk less: compact output without losing commands or errors
- Never forget: session memory carries useful context forward
While you work
Watch the loop tighten in Claude Code
Screen recording of a Claude Code session using Atelier: the agent reaches the right code in fewer turns, and the statusline shows turns, tool calls, and cost dropping live as it works.
Why it's faster
One-shot search cuts the round-trips.
A tighter loop finishes faster.
No prompt tricks. 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.”
Why it works
Not a feature bolted on — a runtime the agent runs inside.
A bare MCP server is a library: tools the model can call if it remembers to. Atelier owns the tool surface — one-shot search in front of the slower native tools, plus role boundaries, procedures, and interrupts — the way an OS owns syscalls.
Find things in one shot — no wandering the codebase call after call.
ReplacesGrep, Glob
WhyOne call returns the symbol, its callers/callees, and ranked source — no grep-loop-then-read-whole-file. Ranked by call-graph centrality over a tree-sitter symbol table.
ReplacesRead
WhyReturns an outline or the exact :L10-L40 range, budgeted, instead of the full file — source projection: outline, compact, or minified views instead of raw text.
ReplacesEdit, Write
WhyVerified, cross-file edits in one call instead of per-file patch-or-create guessing — diff-match-patch under the hood, checked by post-edit contract verification.
ReplacesBash
WhyOutput is capped and structured so a noisy build log can't blow the context window — overflow lands in a recoverable spill store, repeat calls collapse via run-and-dedup.
ReplacesWebFetch
WhyStrips a page to clean Markdown instead of a raw HTML dump — a type-aware summary ladder gists it for free when you only need the gist.
Without itA general-purpose agent edits mid-exploration — nothing stops it.
What closes itexplore / plan / research / review hard-deny edits at the host level.
Without itMulti-step procedures get re-improvised every session.
What closes itSkills encode the procedure once; every agent runs it the same way.
Without itWasteful re-reads and “done” without verification go unchecked.
What closes itHooks deny bad calls before they run and block close until checks ran.
Without itUnder pressure, agents fall back to grep-and-read-whole-file.
What closes itcode_search / read / edit / bash are the only tools visible for those jobs.
Full agent roster, packaged skills, and internals: README · Architecture docs
Measured on public benchmarks
Fewer turns, less wall-clock -- from raw runs, not live badges.
One-shot search means fewer round-trips -- the lower cost falls out for free. Both arms ran the same model, tasks, and environment. Every raw run is committed to the repo.
Scroll to see all columns →
| Benchmark | Baseline | Atelier | Δ correct | Speed | Baseline $ | Atelier $ | Δ cost | Savings |
|---|---|---|---|---|---|---|---|---|
SWE-bench Verified 50 tasks × 5 reps | 80.8% | 92.8% | +12.0 pp | 37.7% fewer turns · 23.7% faster | $234.84 | $165.45 | 29.5% cheaper | |
SWE-bench Lite 10 tasks × 3 reps | 93.3% | 100% | +6.7 pp | 15.8% fewer turns · 5.9% faster | $12.38 | $10.79 | 12.9% cheaper | |
SWE-bench Pro 10 tasks × 5 reps | 88.0% | 90.0% | +2.0 pp | 28.1% fewer turns · 17.3% faster | $39.01 | $30.61 | 21.5% cheaper | |
Exploration 7 large repos × 5 reps | — | — | — | 91% fewer turns | $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 short-answer prompts · 20 × 5 reps | — | — | — | 49.6% fewer turns · 26.3% faster | $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 →
Scroll to see all columns →
| 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
Count the round-trips you are wasting.
Right after install, run the read-only report against your own Claude/Codex history. It counts the wasted tool calls and round-trips in your local files — and the tokens and cost that fall out of them.
$ 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 put a big number on a landing page. 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. Atelier's edge is speed, but Terminal-Bench 2.1 is flat on accuracy (-0.2pp) and only cheaper on cost.
Read the raw runs on GitHub →Not sold yet? Good.
You don't have to install to start a conversation.
Two low-commitment ways in — one human, one just your inbox. No drip sequence either way.
Send me your repo and a task. I'll run Atelier on it and show you the exact savings — live.
A real person reads it and replies with the numbers on your actual code. Not a bot, not a sales call.
Next up: atelier optimize — it finds the savings you're still leaving on the table.
An advisor that reads your own sessions, proposes the exact config change, and shows the measured delta before you apply it. Early access opens in small batches.
Verify first. Install second.
Start the next session with proof, not promises.
Every raw run is published — nothing held back.
Need it for a team? Large codebases, shared team memory, a faster agentic harness at scale.
Contact us about Enterprise