tokentoll

Catch LLM cost changes in code review. Infracost for LLM spend.

A CLI tool and GitHub Action that statically analyzes your code for LLM API calls, estimates their cost, and shows you the cost impact of every change in your terminal or as a PR comment. Zero runtime dependencies.

<p align="center"> <img src="demo/demo.gif" alt="tokentoll demo" width="720"> </p>

The Problem

A single model swap from gpt-4o-mini to gpt-4o increases costs 15x. A new API call in a hot path can add $10,000/month to your bill. These changes hide in normal code review.

tokentoll finds LLM API calls in your code, estimates their cost, and shows you the cost impact of every change before it hits production.

Quick Start

pip install tokentoll

# Scan current directory for LLM API calls and their costs
tokentoll scan .

# Show cost impact of your last commit
tokentoll diff HEAD~1

# Compare two branches
tokentoll diff main..feature-branch

GitHub Action

name: LLM Cost Diff
on:
  pull_request:
    paths:
      - "**.py"

permissions:
  pull-requests: write

jobs:
  cost-diff:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0

      - uses: Jwrede/tokentoll@v0.6.1

What It Detects

SDK	Patterns	Status
OpenAI	chat.completions.create, responses.create	Supported
Anthropic	messages.create, messages.stream	Supported
Google GenAI	models.generate_content	Supported
LiteLLM	completion, acompletion	Supported
LangChain	ChatOpenAI, ChatAnthropic, init_chat_model	Supported
Zhipu AI	ZhipuAiClient, ZhipuAI (GLM models)	Supported
JS/TS SDKs		Planned

Example Output

tokentoll scan

LLM API Calls Detected
============================================================

File: src/agents/summarizer.py
  Line 42: openai client.chat.completions.create
           Model: gpt-4o | Max tokens: 4096
           Est. cost/call: $0.03 | Monthly (1000 calls/month per call site): $26.50

  Line 78: openai client.chat.completions.create
           Model: gpt-4o-mini | Max tokens: 1000
           Est. cost/call: $0.000301 | Monthly (1000 calls/month per call site): $0.30

--
Total estimated monthly cost: $26.80
  1000 calls/month per call site

tokentoll diff

LLM Cost Diff: main..feature-branch
============================================================

+ ADDED    src/agents/rewriter.py:35
           openai | Model: gpt-4o
           Est. cost/call: $0.03 | Monthly: +$26.50

~ MODIFIED src/agents/summarizer.py:42
           openai | Model: gpt-4o -> gpt-4o-mini
           Est. cost/call: $0.03 -> $0.000301 | Monthly: -$26.20

--
Monthly cost impact: +$0.30
  Added: 1 | Changed: 1 | Removed: 0
  1000 calls/month per call site

How It Works

  Source Code (.py files)
         |
         v
  +-------------+     +------------------+
  | AST Scanner |---->| SDK Detectors    |
  | (ast.parse) |     | OpenAI, Anthropic|
  +-------------+     | Google, LiteLLM  |
                       | LangChain        |
                       +------------------+
                              |
                              v
                       +------------------+
                       | Pricing Engine   |
                       | 2200+ models     |
                       | Auto-cached      |
                       +------------------+
                              |
                  +-----------+-----------+
                  |                       |
                  v                       v
           +------------+         +-------------+
           | Scan Report|         | Diff Engine  |
           | (costs)    |         | (old vs new) |
           +------------+         +-------------+
                  |                       |
                  v                       v
           +------------+         +-------------+
           | Table/JSON |         | Table/JSON/  |
           |            |         | PR Comment   |
           +------------+         +-------------+

1. Parses Python files using the ast module to find LLM API calls
2. Multi-pass constant propagation resolves model names through variables, os.getenv() fallbacks, class attributes, constructor args, dict contents, and **kwargs unpacking
3. Looks up pricing from a local cache (sourced from LiteLLM, 2200+ models)
4. For diff mode: compares calls between two git refs and computes the cost delta
5. Outputs a cost report as a table, JSON, or GitHub PR comment

CLI Reference

tokentoll scan [PATH...] [--format table|json|markdown] [--calls-per-month N] [--config PATH]
tokentoll diff [REF] [--base REF] [--head REF] [--format table|json|markdown|github-comment] [--config PATH]
tokentoll update    # Update bundled pricing data

MCP Server

tokentoll includes an MCP (Model Context Protocol) server that lets Claude Code and other MCP hosts check the cost impact of LLM code changes directly from an agent conversation.

Install

pip install tokentoll[mcp]

Register with Claude Code

claude mcp add --transport stdio tokentoll -- tokentoll-mcp

Tools

Tool	Description
scan	Find LLM API calls in a directory and estimate monthly costs. Accepts a path and optional calls_per_month.
diff	Compare LLM costs between two git refs. Accepts base_ref and optional head_ref (defaults to HEAD).

Both tools return JSON output.

Example use case

Claude Code can check the cost impact of its own changes before committing. For example, after swapping a model from gpt-4o to gpt-4o-mini, the agent can call the diff tool against HEAD to verify the cost reduction before creating the commit.

Pricing Data

Pricing is bundled and works offline. To update to the latest prices:

tokentoll update

Pricing data is sourced from LiteLLM's model_prices_and_context_window.json and covers 300+ models across OpenAI, Anthropic, Google, AWS Bedrock, Azure, and more.

Dynamic Model Defaults

When tokentoll encounters a call where the model name is a variable it cannot resolve, it applies a sensible per-SDK default so you still get cost estimates:

SDK	Default Model
OpenAI	gpt-4o
Anthropic	claude-sonnet-4-20250514
Google GenAI	gemini-2.0-flash
LiteLLM	gpt-4o
LangChain	gpt-4o
Zhipu AI	zai/glm-4.6

These defaults are shown as gpt-4o (default) in scan output. You can override them per-project or per-path using a .tokentoll.yml config file (see below).

Configuration

Create a .tokentoll.yml in your project root to customize behavior. tokentoll automatically finds this file by walking up from the scanned directory.

# Default model for all dynamic (unresolved) calls
default_model: gpt-4o

# Per-SDK defaults (override the built-in defaults above)
default_models:
  openai: gpt-4o-mini
  anthropic: claude-haiku-3-20240307

# Assumed calls per month per call site
calls_per_month: 5000

# Skip cost estimation entirely for dynamic (unresolved) models. When true,
# calls whose model name cannot be resolved statically are reported with no
# cost rather than priced against a default. Useful for projects that prefer
# silence over a guess.
skip_dynamic_models: false

# Exclude paths from scanning (prefix match or glob pattern)
exclude:
  - tests/
  - examples/
  - docs/
  - "*_test.py"

# Per-path overrides (longest prefix match)
overrides:
  - path: src/agents/
    default_model: gpt-4o
    calls_per_month: 10000
  - path: src/azure/
    skip_dynamic_models: true

Resolution order for dynamic model defaults: per-SDK config (default_models) > generic config (default_model) > built-in SDK defaults.

You can also pass --config path/to/.tokentoll.yml to use a specific config file.

Token Estimation

By default, tokentoll estimates token counts using a characters/4 heuristic. For more accurate estimates, install tiktoken:

pip install tiktoken

When tiktoken is available, tokentoll uses the correct tokenizer encoding for each model. Unknown models fall back to cl100k_base. Tiktoken is lazy-loaded and encoders are cached, so there is no startup penalty if you don't need it.

Smart Variable Resolution

Real codebases rarely pass model names as string literals. tokentoll's multi-pass constant propagation engine follows:

DEFAULT_MODEL = os.getenv("MODEL", "gpt-4o")

class Config:
    model: str = DEFAULT_MODEL

config = Config()
kwargs = {"model": config.model, "max_tokens": 2000}
client.chat.completions.create(**kwargs)
# tokentoll resolves: model="gpt-4o", max_tokens=2000

• Variable assignments (MODEL = "gpt-4o")
• os.getenv() / os.environ.get() fallback values
• Function default parameters
• Class attribute defaults
• Constructor argument propagation
• Dict literal and subscript contents
• **kwargs unpacking

Roadmap

• Context-aware call frequency (planned): infer calls/month from surrounding code (FastAPI route handlers = high traffic, scripts = low, loops = multiplied) instead of assuming uniform volume across all call sites.
• JS/TS support (planned): detect LLM calls in JavaScript and TypeScript files.
• Cost alerts: configurable thresholds that fail CI when a PR exceeds a cost delta.

Limitations

• Cannot resolve models loaded from external config files or databases at runtime. These calls use per-SDK defaults (configurable via .tokentoll.yml).
• Token estimates use a characters/4 heuristic unless tiktoken is installed.
• Monthly estimates assume uniform call volume per call site (configurable via --calls-per-month, .tokentoll.yml, or per-path overrides). Use the exclude option to skip test and example files.
• Python only for now (JS/TS support planned).

License

MIT