The mirrorkit collectors batch traces and POST them here in the background. If you have your own pipeline — an existing observability platform, a custom framework, a language without a collector — you can speak the wire format directly.
Request
POST https://api.runmirrors.com/api/collect
Authorization: Bearer mk_live_...
Content-Type: application/json
{
"project": "my-agent",
"traces": [
{
"id": "abc",
"model": "gpt-4o",
"messages": [
{ "role": "user", "content": "What's the weather in Paris?" },
{
"role": "assistant",
"content": null,
"tool_calls": [
{ "id": "call_1", "function": { "name": "get_weather", "arguments": "{\"city\": \"Paris\"}" } }
]
},
{ "role": "tool", "tool_call_id": "call_1", "content": "18C, sunny" },
{ "role": "assistant", "content": "It's 18C and sunny in Paris." }
]
}
]
}
Fields
The project the traces belong to. Mirrors are built per project, so keep one project per agent.
A batch of traces. Each trace has an id (string, unique per trace), an optional model, and messages — an OpenAI-style chat array (system / user / assistant / tool roles, with tool_calls on assistant messages and tool_call_id on tool results).
Behavior
- Batching is up to you; the official collectors default to 50 traces per POST, flushed every 2 seconds.
- Requests are authenticated by workspace key; traces land in that workspace under
project.
- The collectors retry failed batches (non-2xx / network errors) a couple of times, then drop them — ingest is designed to never break the instrumented app.
Tool calls and their results are the most valuable part of a trace — they’re what Mirrors uses to infer your agent’s world. Include them whenever you can.
