Why we never see your prompts and how that works

The architecture

Tokeven's wrapper runs in your environment - your machine, your server, your container. It sits between your code and the provider SDK. When a call completes, it reads the usage metadata from the response and pushes a small event to Tokeven's ingest API.

What we receive

• -Model ID (e.g., claude-sonnet-4-6)
• -Token counts (input, output, cache-read, cache-write)
• -Latency in milliseconds
• -Provider, modality and a local prompt character count (never the text)
• -An optional session identifier and project label (for grouping related calls)
• -An optional actor label naming who ran the call (off by default; if you set it to an email address, that address is sent)

The full list is in our Privacy Policy.

What we never receive

• -Prompt content
• -Completion content
• -API keys
• -File contents or code

Why this matters

Most analytics tools ask you to proxy your traffic through their servers. That means they see everything - your prompts, your completions, your keys. Tokeven doesn't work that way.

The wrapper runs locally. It reads the response object's usage field after the call completes. The prompt text goes to your AI provider as it always did, and is never sent to or stored by Tokeven. (A few opt-in tools outside this pipeline do process text you explicitly submit, and they are listed in Privacy Policy §2.4.)

The ingest token

Authentication to Tokeven uses a write-only ingest token (tkv_ing_...). It can only push metrics - it can't read your data, can't access your account settings, and can't be used to make API calls to any provider. If it leaks, the worst case is someone pushes fake metrics to your dashboard.

Verification

The wrapper ships as a Python package (`tokeven` on PyPI) under a proprietary licence, and the installed source is readable, so you can inspect exactly what it captures and transmits. `tokeven status` reports where events are being sent and how many are queued. The ingest API validates the schema and rejects any field it does not recognise, so there is nowhere for content to go.