Tool calls used to be buried inside raw response JSON. Now every tool your agents invoke is extracted into a dedicated table and ready to query.
Per-tool stats — Call volume, p50/p95 latency, error rate, and orphan count (tool calls that never got a result) for every tool your agents use. Sortable, filterable, rolled up across any time range.
Volume over time — A stacked chart of tool call volume broken down by the tools you're running the most. Spot a runaway retry loop or a tool being hit more than you expected at a glance.
Historical coverage — We backfilled your existing traffic, so the analytics reach all the way back. No waiting for new data to accumulate.
Every provider, every shape — OpenAI Chat Completions, the OpenAI Responses API, Anthropic Messages, and streaming (SSE) variants of each. One view, all providers.
Find it under Analytics → Tools.
What's next
This is the foundation. With tool calls in a real table, we can reason about the whole agent loop — not just individual LLM calls. On the roadmap:
- Agent runs as a first-class unit — group LLM calls and tool calls into logical runs with step counts, terminal state, and cost rollups.
- Tool-layer safety checks — detect credential-shaped arguments heading to external APIs, and flag prompt-injection-style content inside tool results before your agent acts on it.
More soon.