TokenSentinel Documentation
Welcome to the official documentation for TokenSentinel — predictive token-waste detection and intervention for AI agents.
| Package | Version / status | Notes |
|---|---|---|
Python SDK (token-sentinel) | 1.0.3 | Open source (Apache-2.0). 15 rules, model-aware burn estimates, optional tiktoken. |
| TokenSentinel Cloud | Coming Soon | Dashboards, Intervention Pack, Pro features and more. For more details, visit official website tokensentinel.dev |
The Python SDK runs in-process on your LLM clients (15 deterministic rules, 9 native provider families). Detection runs after each provider response; you can log, alert, or hard-stop the agent before the next call. Optional TokenSentinel Cloud (paid) adds dashboards, the Intervention Pack, and Pro calibration.
Site: docs.tokensentinel.dev · Web: tokensentinel.dev · SDK: GitHub · PyPI
pip install "token-sentinel>=1.0.3,<2"
python -c "import token_sentinel; print(token_sentinel.__version__)" # -> 1.0.3
Documentation sections
Python SDK Reference — SDK 1.0.3
Install, wrap clients (Anthropic, OpenAI + compatible, Gemini, Bedrock, Voyage, Cohere, Replicate, Deepgram, ElevenLabs), configure all 15 rules, modes, integrations (LangChain / OTel), and the public API.
Release notes: CHANGELOG on GitHub.
Migration Tool — migrate 0.1.2
Replay historical traces from Helicone, Langfuse, or LangSmith through the SDK rules engine (15 rules with current token-sentinel) and optionally backfill to cloud.
SaaS user journey (coming soon)
Hosted-dashboard onboarding, budgets, alerts, and Pro features are coming soon, For more details, visit official website tokensentinel.dev
Support: support@tokensentinel.dev · Hello: hello@tokensentinel.dev