How it works
What waste looks like—and how you wire it in.
The homepage covers why mid-run matters. Here: the patterns we catch, and where TokenSentinel sits in your stack.
Patterns
Examples of waste we help you catch.
- Agents stuck repeating the same tool
- Retry storms after provider errors
- Context that grows without bound
- Silent sessions still spending tokens
- Expensive models used for trivial tasks
- Repeated retrieval of the same work
In your stack
Wrap the client. React to the signal.
In-process—not a proxy. Complements tracing tools.
One wrap around the model client you already use. Works across major providers.
- Anthropic
- OpenAI
- Google Gemini
- AWS Bedrock
- Cohere
- Voyage
- OpenAI-compatible endpoints
Setup: get started · docs
Also
Claude Code plugin · beta
Optional plugin for Anthropic’s Claude Code. Watches local coding sessions for stuck tool loops and similar waste. Quiet by default.