Which provider breaks your agent's tool calls? Daily data for every open model.
Tool calls that work on every model.
Every model ships its own tool-call format, and every engine needs its own parser. .lambda writes each model's format once and ships it to vLLM, SGLang, TensorRT-LLM and ONNX Runtime GenAI from the day the model is released, with a per-token cost bound in the contract. It runs inside your stack, and nothing leaves your cluster.
Same model, very different tool calls depending on who serves it.
Routers now rank provider endpoints on how often their tool calls fail, and send the traffic to the top. On the busiest open-weight models, the same model fails a fraction of a percent of its tool calls on its best provider and several percent on its worst. The tool-call index publishes those rates for every provider, every day, at providers.dottxt.ai.
Day 0
With the model's weights, chat template and tokenizer before release, its tool-call format is validated before launch. Otherwise it ships inside a contractual window.
The same answer on every engine
Whether a tool call is valid no longer depends on which engine served it. Move a model from vLLM to TensorRT-LLM without re-validating its tool calls.
Enforced or rejected
A schema is enforced exactly, or the request fails and names the keyword it can't enforce. Never a silently loosened result returned with a 200.
Your fallback stays
One patch per runtime, behind the open-source engine's interface, with versioned releases and rollback. No calls home.
What you sign for
A per-token bound
The cost per token is bounded in the contract, so the slowest lane never sets the pace of the batch.
Formats, maintained
We write and test each new model's tool-call format. Fixes to models you already serve never count against your release budget.
Streaming
Tool calls are checked while they are generated, so arguments stream as they are produced.
An SLA by severity
An incident gets a severity and a response time, not an issue on a tracker.