tokenlint 0.1.0

Catch LLM cost anti-patterns in your Python AI app before they hit prod — a local, static token-cost linter.

TokenLint

Your LLM bill is bleeding and you don't know which line of code is doing it. TokenLint scans your Python AI app locally and statically (it never runs your code or sends it anywhere) and flags the patterns that quietly inflate token costs — premium models on trivial steps, missing output caps, un-cached static prompts, and unbounded agent loops that burn tokens unpredictably — before they hit production.

Runs in your editor, your terminal, or CI. Zero dependencies, zero network calls.

pipx install tokenlint     # or: pip install tokenlint
tokenlint .                # scan your project

TokenLint  — LLM cost pre-flight

  app/agent.py
    FAIL R004-unbounded-agent-loop  line 30
         LLM call inside a `while True` with no break — token runaway risk.
         → Add an explicit max-steps budget and a token/cost ceiling.
         ~$0.0050/call (gpt-4o)
    WARN R001-expensive-model  line 19
         Call uses 'claude-opus-4', a premium-priced model.
         → If this step is simple, route it to claude-sonnet/haiku.
         ~$0.0377/call

  1 fail   7 warn   0 info

What it catches today (v0.1, free, MIT)

Rule	What it flags
R001-expensive-model	Premium model (Opus / GPT-4o / o1…) on a step that a cheaper model could do
R002-no-max-tokens	No output cap — one runaway completion can 10× a call's cost
R003-no-prompt-cache	Large static Anthropic system/context with no cache_control (often ~90% of input spend is cacheable)
R004-unbounded-agent-loop	LLM call inside while True with no break — the #1 way agents quietly burn money
R004-llm-call-in-loop	LLM call in a loop — cost scales with iterations; needs a budget
R005-large-inline-prompt	Oversized inline prompt literal sitting in the hot path

Works with the Anthropic and OpenAI SDKs and raw call patterns. LangChain / LlamaIndex detection is on the roadmap.

CI usage

tokenlint . --fail-on warn   # exit non-zero on warnings too, to block a costly merge

Use it inside Claude Code (plugin)

TokenLint ships as a Claude Code plugin: a /tokenlint:cost-preflight skill that scans your project and offers fixes, plus a non-blocking hook that flags an unbounded agent loop the moment you write one.

# from the repo (once it's on GitHub)
/plugin marketplace add zlyan1110/tokenlint
/plugin install tokenlint@tokenlint

# or try it locally without installing
claude --plugin-dir ./tokenlint

Then in a session:

/tokenlint:cost-preflight          # scan the whole project
/tokenlint:cost-preflight app/     # scan a path

The skill auto-loads when you're working in .py files and will offer to apply fixes (caps, prompt caching, model routing, agent budgets) with your approval.

TokenLint Pro

The free linter tells you where you're leaking. Pro fixes it for you:

• Auto-fix diffs — PR-ready patches that add caps, prompt caching, and budgets
• Model-routing config generator — send cheap steps to cheap models automatically
• Agent token-budget guard — a drop-in wrapper that hard-stops runaway loops
• Self-host break-even calculator + vLLM/TGI deploy configs — when does running your own inference beat the API, and how to configure it

→ https://tokenlint.dev/pro · $39 one-time or $19/mo (always-current pricing + rules)

Why local + static?

Observability platforms (Helicone, Langfuse, …) show you cost after traffic runs through them. TokenLint catches it before you ship — in the same loop where you write the code (including inside Claude Code / Cursor). The two are complementary; this one is free and runs offline.

License

MIT for the core linter. Pricing table in tokenlint/pricing.py is editable — verify against current provider pricing.