Trace Audit — a diagnosis without an API key
Failed-task costs, repeated calls and evidence: free analysis of 1,000 events, directly available through API or MCP.
Try it now
Open /en/tools/trace-audit. The complete example is computed immediately in your browser. Replace it with your events to get your diagnosis. JSON downloads require no account.
One API call, no signup
curl https://alpnai.com/api/v1/trace-audit \
-H 'Content-Type: application/json' \
--data-binary @events.jsonEvent format
events contains one row per call: unique event_id, task_id, tool and numeric cost_usd with up to 6 decimal places. input_hash is an opaque fingerprint computed by you; do not send arguments, prompts or keys. latency_ms is optional.
Results and measurement scope
outcomes contains at most one verdict per whole task: success, failed or unknown. Omitted tasks remain unknown. Cost per success includes ALL batch costs, divided by the number of tasks explicitly labelled successful. With no successes it returns null. Repetitions are candidates to review, not proven savings; their cost can already be included in failed-task costs.
Connect a compatible MCP client
The public server offers audit_agent_trace and get_trace_example without a key or payment. Add this URL as a Streamable HTTP server with no authentication in a compatible client. In ChatGPT, custom connection availability depends on your account plan and settings. ALPNAI does not claim a published ChatGPT directory app. The existing /api/mcp server keeps its private tools and authentication.
https://alpnai.com/api/mcp/publicData handling
Maximum 1,000 events and 256,000 HTTP request bytes. Local computation keeps events in your tab only. The API/MCP server processes the batch without storing it or calling an external model. Only aggregate operational counters are recorded. No orders, payments or agent changes are made.