Tokentoll

What's New

tiktoken support

Token estimation now uses tiktoken when installed, replacing the char/4 heuristic. Install with pip install tokentoll[tiktoken].

Project config file (.tokentoll.yml)

Projects can add a .tokentoll.yml to configure default models for dynamic call sites:

default_model: gpt-4o
calls_per_month: 5000

overrides:
  - path: src/agents/
    default_model: claude-sonnet-4-20250514
  - path: src/embeddings/
    default_model: text-embedding-3-small

When tokentoll cannot statically resolve a model name, it uses the configured default instead of showing <dynamic> | N/A. Output shows gpt-4o (default) to distinguish from resolved models.

This makes tokentoll useful for router/proxy projects where model names are always determined at runtime.

CLI changes

• New --config flag on scan and diff to specify a config file path
• --calls-per-month now defaults to config value when not explicitly set

71 tests, all passing.