cd /news/ai-tools/i-was-tired-of-all-false-claims-so-i… · home topics ai-tools article
[ARTICLE · art-51948] src=github.com ↗ pub= topic=ai-tools verified=true sentiment=↑ positive

I was tired of all false claims, so I built Atelier, honest 30% saving on CC.

Atelier, a 30-second install tool that sits underneath Claude Code to reduce token waste, claims 30% savings on AI coding costs by providing better search, shorter file reads, compact command output, and reusable memory. The tool offers a live savings meter and allows users to verify savings by scanning their own local agent history, with benchmark results showing cost reductions of 12.9% to 67% across various tasks.

read5 min views1 publishedJul 9, 2026
I was tired of all false claims, so I built Atelier, honest 30% saving on CC.
Image: source

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

Run this once:

curl -fsSL https://install.atelier.ws | bash

Then turn it on inside the project where you use Claude Code:

cd your-project
atelier init

Open Claude Code like you normally do. Atelier wires in better tools behind the scenes and starts tracking savings as sessions finish.

Already installed?

atelier update

Check that everything is connected:

atelier doctor

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.

Do not take our 30% claim on faith. Before installing, you can scan your own local agent history:

curl -fsSL https://savings.atelier.ws | bash

What 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

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.

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/.

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.

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.

  • 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.

InstallationTroubleshootingBenchmarks·full results, backed by docs·every "vs" comparison, with sourcesCLI referenceArchitecture

Apache 2.0 — see LICENSE.

── more in #ai-tools 4 stories · sorted by recency
── more on @atelier 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/i-was-tired-of-all-f…] indexed:0 read:5min 2026-07-09 ·