Honestly, get 30% more out of your Claude subscription
Atelier is a 30-second install that helps Claude Code waste fewer tokens while you work. Keep using Claude Code normally; Atelier sits underneath it and gives the agent better search, shorter file reads, compact command output, reusable memory, and a live savings meter.
Live savings across Atelier sessions
Install · Check your savings first · Why trust the numbers? · Results · Pricing
Install in 30 seconds
Run this once:
curl -fsSL https://install.atelier.ws | bashThen turn it on inside the project where you use Claude Code:
cd your-project
atelier initOpen Claude Code like you normally do. Atelier wires in better tools behind the scenes and starts tracking savings as sessions finish.
Already installed?
Check that everything is connected:
What changes for you
Atelier does not ask you to learn a new coding app. It improves the work Claude Code already does:
| Before | With Atelier |
|---|---|
| Claude reads broad files and long terminal output. | Claude gets the exact code ranges and compact results it needs. |
| The same context gets rediscovered again and again. | Useful session context can be reused. |
| You pay for long explanations inside the working transcript. | Outputs stay shorter while code, commands, filenames, and errors remain exact. |
| Savings are hard to see. | A local meter shows tokens, cost, and savings adding up. |
The point is simple: more of your Claude subscription should go into useful work, not repeated setup and paid noise.
Check your own savings
Do not take our 30% claim on faith. Before installing, you can scan your own local agent history:
curl -fsSL https://savings.atelier.ws | bashWhat it does:
- Reads local Claude/Codex agent session files.
- Estimates where Atelier would have used fewer tokens or cheaper tool calls.
- Prints savings from your own history.
- Uses a temporary local store.
- Needs no Atelier account and no provider API keys.
Useful variants:
curl -fsSL https://savings.atelier.ws | bash -s -- --since 30d --top 10 curl -fsSL https://savings.atelier.ws | bash -s -- --host codex --limit 20
Why trust the numbers?
You are right to be skeptical. A live badge alone proves very little because anyone can fake a counter.
Atelier uses four checks instead:
- Raw benchmark receipts: headline numbers link to committed per-task runs, costs, turn counts, setup notes, and reproduction commands.
- Your own scan: the savings command checks your machine, not our marketing page.
- Labeled live badges: live counters show aggregate usage adding up; they are not used as the source of the 30% benchmark claim.
- Rows where Atelier does not win: Terminal-Bench 2.1 is flat on accuracy (-0.2pp) and only cheaper on cost. It stays in the results table.
The trust is the audit trail, not the animation.
Results
Measured on the same model, same tasks, and same environment:
| Benchmark | Baseline correct | Atelier correct | Correct delta | Baseline cost | Atelier cost | Cost delta |
|---|---|---|---|---|---|---|
| SWE-bench Verified, 50 tasks x 5 reps | 80.8% | 92.8% | +12.0 pp | $234.84 | $165.45 | 29.5% cheaper |
| SWE-bench Lite, 10 tasks x 3 reps | 93.3% | 100% | +6.7 pp | $12.38 | $10.79 | 12.9% cheaper |
| SWE-bench Pro, 10 tasks x 5 reps | 88.0% | 90.0% | +2.0 pp | $39.01 | $30.61 | 21.5% cheaper |
| Exploration tasks across 7 large repos x 5 reps | - | - | - | $19.11 | $6.29 | 67% cheaper |
| Telegraphic Q&A, 20 prompts x 5 reps | - | - | - | $8.93 | $5.34 | 40.2% cheaper |
| Terminal-Bench 2.1, 89 tasks vs public leaderboard* | 78.9% expected | 78.7% | -0.2 pp | $96.76 | $69.52† | 28.1% cheaper† |
* Atelier: 1 rep/task. Baseline: public tbench.ai leaderboard, 5-rep average per task. † Other 5 tasks in Atelier timeout and cannot capture cost; see .
SWE-bench Verified detail:
| Metric | Baseline | Atelier | Delta |
|---|---|---|---|
| Turns | 6,962 | 4,336 | 37.7% fewer |
| Output tokens | 3.04M | 2.19M | 27.9% fewer |
| Wall-clock | 14.3h | 10.9h | 23.7% faster |
Raw runs, setup notes, and reproduction commands live in BENCHMARKS.md and benchmarks/codebench/results/.
Code search vs 10 named tools
Retrieval quality (MRR) across ~7,200 query/gold pairs on 14 repos -- ripgrep, ast-grep, universal-ctags, Serena, CodeGraph, cocoindex-code, codebase-memory-mcp, fff-mcp, code-index-mcp, and jCodeMunch all scored on the identical corpus:
| Provider | MRR | rec@1 | p95 |
|---|---|---|---|
| Atelier +semantic (BGE) | 0.727 | 0.650 | 390ms |
| Atelier lexical (default) | 0.676 | 0.582 | 134ms |
| cocoindex-code (best rival) | 0.557 | 0.457 | 595ms |
| serena | 0.401 | 0.359 | 3834ms |
| ripgrep | 0.376 | 0.320 | 66ms |
| universal-ctags (worst rival) | 0.237 | 0.226 | 1ms |
No one had scored these 10 tools against each other on a shared query set before -- each publishes its own number, on its own terms, against its own baseline. Full 13-row table and per-tool "claim vs. scored" breakdown: atelier.ws/vs · docs.atelier.ws/benchmarks/results.
Also worth a look, two comparisons outside code search: rtk, a 69.6k-star Rust CLI proxy -- self-estimated savings vs. Atelier's accuracy-checked Terminal-Bench number. "Just tell Claude to be terse", the free DIY alternative -- benchmarked head-to-head on the same 20 prompts, including the one where it backfires.
Why it works
Claude is strong, but the work around Claude is often wasteful. Atelier reduces that waste.
- Better inputs: the agent gets relevant symbols and exact file ranges instead of whole files.
- Better outputs: command output and replies stay compact without losing exact technical facts.
- Better memory: useful context can be reused instead of rediscovered.
- Better guardrails: tools and hooks reduce risky edits, oversized reads, and unverified "done" states.
Atelier does not make Claude a different model. It makes the loop around Claude cleaner, which is why the same model solved more tasks in the benchmark.
What you get
- Works with Claude Code, Codex, Copilot, Cursor, opencode, Hermes Agent, LangChain, the OpenAI SDK, Gemini ADK, and any other MCP-compatible coding agent.
- Runs locally by default.
- Apache-2.0 open source runtime.
- No account needed to start.
- Live local stats for cost, tokens, and savings.
- Optional paid features for heavy users and teams.
Learn more
- Installation
- Troubleshooting
- Benchmarks · full results, backed by docs · every "vs" comparison, with sources
- CLI reference
- Architecture
License
Apache 2.0 — see LICENSE.