Skip to main content
Open Observe → Agents to see the agents and model families recorded in your workspace. The Agents and Models tabs show calls from the selected 7 or 30 days, clipped to your plan’s retention. Open a card for usage, cost, call errors, latency, related identities, recorded tools and recent traces.

​
Name an agent

For a request through /v1/chat/completions, /v1/embeddings or /v1/messages, send this header:
The gateway trims and lowercases the name, truncates it to 64 characters, and ignores empty values or values containing control characters. Use a project or agent name rather than personal information. The header does not affect access control or workspace selection. For SDK and OpenTelemetry capture, set the standard attributes on the LLM span:
Use embeddings or messages for those operations. Spans without one of these operation names do not count as Calls, even when they carry an agent name. Some SDK integrations require you to supply the operation attribute explicitly. A supplied agent name takes priority over a recognised client. The gateway’s client classifier stores only a catalog name, such as claude-code, or nothing for an unknown client. It does not store the request’s User-Agent in the span. A recognised client name describes the request; it does not establish who sent it. The currently observed Codex CLI uses the Responses API. This gateway does not serve /v1/responses; a catalog entry alone does not make that protocol work. Use a supported gateway route or SDK span capture for identity activity.

​
Read a card

  • Direct API calls includes calls without an agent name or recognised client.
  • Unidentified model includes calls without a model name. Neither bucket is dropped.
  • Response model takes priority over request model. Route prefixes and dated suffixes are removed to group model families.
  • Made by comes from the catalog; Served by comes from the recorded provider. A Meta model served by Groq shows both.
  • An identity outside the catalog gets its recorded key and a monogram.
  • Identity source distinguishes SDK names, header names, recognised clients and unnamed calls. These are caller-supplied signals.
Calls counts only chat, embeddings and messages spans, not tool, agent or internal spans. Tokens and cost show a dash when all calls lack that measurement; calls missing usage or prices are counted separately. First and last seen are within the selected window. Error rate counts spans with error status; latency uses recorded span duration. Tools are names captured on tool spans or model responses. Without captured tool names, the profile says No tool calls recorded. The directory shows at most 200 identities by Calls; related identities and tools show the top 10 with a remaining count. Profiles link to 20 recent traces. View all in Traces keeps the identity and selected time window as filters, including exports.

​
Read activity through the API

Use a session token or an API key with the read scope. The gateway obtains the workspace from that credential; there is no caller-supplied tenant parameter.
Use kind=model for model families. Identities with no calls in the selected window and identities outside your workspace return the same 404 on the profile route. Failed reads remain errors, rather than turning into zero activity.