Privacy

Prompts and completions are never transmitted. This page lists exactly what is.

Transmitted

Per call:

• model and provider
• token counts: input, output, cache read, cache write
• computed cost
• latency
• attribution: agent, team, task type, session, customer, cost centre
• environment label
• the file and line that made the call
• whether the call was routed, and from which model
• whether it failed, and the exception's type name

Never transmitted

• prompt text, system prompts, message content
• completion text
• tool definitions, tool arguments, tool results
• embedding inputs and output vectors
• exception messages, because provider errors often quote the prompt back
• your provider API keys, which the SDK does not read

Prompt analysis records structure only

Optional prompt analysis (enable_prompt_analysis=True, off by default) adds structural metrics:

The system-prompt hash lets the backend detect that many calls shared one system prompt, which is how an uncached repeated preamble becomes visible. It is a one-way digest and cannot reproduce the prompt.

Two fields describe your code

caller_file is a path and line number from your repository, such as app/agents/planner.py:44. This is what makes spend traceable to code, and it does describe your project layout.

Agent names are stored as you write them. @capsera.agent("acme-onboarding-flow") is recorded verbatim. Attribution values are the only place your own strings are stored, so avoid names that are themselves sensitive. customer_id is intended for an opaque identifier rather than a customer name.

The canary test

This guarantee is enforced by a test. The evaluation harness plants unique canary strings inside prompts, runs the full corpus, and scans every emitted payload for them. A single occurrence fails the run, alongside phantom events, duplicate billing, and false budget blocks.

A future change that started capturing message content would fail that gate rather than merging quietly.

Self-hosting

Events go to the endpoint you configure. Pointing it at your own deployment keeps everything above inside your infrastructure:

capsera.init(
    api_key=os.environ["CAPSERA_API_KEY"],
    endpoint="https://capsera.internal.example.com",
)

The SDK has one destination, the one you set. There is no secondary telemetry channel and no usage analytics of its own.