🚂 ATrain

Cut your Claude Code bill ~80%. Same accuracy. No API key. 30-second install.

Try first · Install · How it works · Honest numbers

---

What is this

Claude Code runs every tool call on the same model. A 50-character Read pays the same Opus rate as a 4-file refactor.

ATrain is a hook plugin that watches each tool call and routes it to the cheapest Claude model that can handle it. Haiku for reads. Sonnet for edits. Opus only for architecture, security, and anything sensitive.

You write code the same way. Your bill shrinks.

---

Try it on your own past sessions first

No install needed. The autopsy script reads any past Claude Code transcript and tells you exactly what ATrain would have saved.

git clone https://github.com/LeonardoCalancea/atrain-claude
cd atrain-claude
python3 tools/atrain_autopsy.py ~/.claude/projects/*/<session-id>.jsonl

Output on a real 913-prompt session:

🚂 ATrain Token Autopsy
─────────────────────────────────
Prompts analyzed   : 913
Routed to haiku    : 127 (14%)
Routed to sonnet   : 494 (54%)
Routed to opus     : 292 (32%)
─────────────────────────────────
Cost with ATrain   : $12.96
Cost all-Opus      : $44.59
SAVED              : $31.63  (70.9%)

If the number isn't convincing, don't install. If it is, keep reading.

---

Install

git clone https://github.com/LeonardoCalancea/atrain-claude && cd atrain-claude
bash install.sh

Restart Claude Code. Then:

/atrain-ultimate

That's it. Stack armed for the whole conversation. Switch to readable prose with /atrain-regular or disarm with /atrain-kill.

---

Six commands. That's the whole surface.

Command	What it does	Best for
/atrain-ultimate	Max $ savings. Caveman ULTRA + full v8 stack.	API-billed users, $ matters most
/atrain-lean	Min token usage. Routing + caveman + bash-rewrite only. No advisories, no fan-out.	Weekly-limit users (Pro/Max plans)
/atrain-regular	Full v8 stack, caveman OFF, readable prose.	Want explanations, docs, prose
/atrain-kill	Disarm everything. Data retained in cache DB.	Pause
/atrain-status	Live card — cost, savings %, accuracy, tier mix.	Mid-session check
/atrain-autopsy	Project savings on any past transcript.	Try before install

---

How it works

1. Routes every tool call. Recon (Read, Grep, LS) → Haiku. Edits → Sonnet. Architecture, security, large refactors → Opus. Sensitive keywords (auth, payment, crypto, schema migrations, prod deploys) always force Opus, never silently downgraded.
2. Compresses output. Caveman terse mode strips filler from natural-language responses. Code, commits, and security writeups stay normal.
3. Caches and indexes. Repeated Reads served from SQLite. Codebase symbol index built once per session. Cross-session FTS5 recall surfaces prior tool outputs when you ask similar questions.
4. Rewrites bash output. pytest/cargo test/npm test/git status all get compressed before hitting context (Copilot pattern).

Pure stdlib Python. No PyPI dependencies. No API key. No new CLI. Runs inside Claude Code on bundled tokens.

---

Honest numbers

Two real-world measurements. Pick the one that matches your baseline.

Baseline	Saved
Naive Opus xhigh + no output compression	~95%
Same Opus xhigh + same caveman output compression (apples-to-apples)	~80%
Typical Claude Code Sonnet defaults	~40-50%

Across 13 real coding sessions (6,000+ prompts), full-cost accounting including input tokens:

• /atrain average: ~73% saved
• Range: 64–82%
• Classifier accuracy: 108/108

Reproduce yourself:

python3 tools/atrain_full_efficiency_bench.py --stack ultimate
python3 -c "import json, pathlib; print(json.load(open(pathlib.Path.home()/'.claude'/'router-config.json'))['session_stats'])"

---

Drawbacks

• /atrain-ultimate gives terse output. Use /atrain-regular for full prose.
• First session on a new project: no prior cache, savings start lower and ramp up after 2-3 sessions.
• Doesn't help if you're already running everything on Haiku. Helps most against Opus-heavy or default Sonnet workloads.

---

Built on

Patterns credited inline in router.py. Highlights: Skeleton-of-Thought (arxiv 2307.15337), TokenSkip (2502.12067), Adaptive-Consistency (2305.11860), Speculative Cascade (2506.04203), MemGPT (2310.08560), Selective Context (2310.06201), Anthropic's Code-Execution-with-MCP, JuliusBrussee/caveman, rtk-ai/rtk.

---

License

MIT. Use it, fork it, ship it.

---

If it saves you money, star the repo.