For agents
Point any agent at https://docs.assay.solutions/llms.txt. That file is the table of contents. The agent loads only the pages the task needs.
What you can do
- Capture sessions from Claude Code, Grok, Codex, and OpenCode.
- Meter a custom agent loop as it runs.
- Open a local console: spend, sessions, products, objectives, waste, ROI, search.
- Label sittings so spend rolls up to the thing you were building.
- Share one session as a flat page — nothing else is reachable from the link.
- Search what was actually said, without sending the corpus to a model.
What it is not
- Not a cloud dashboard that ships your prompts away.
- Not a proxy. Assay never sits between you and the model.
- Not a value oracle. It records labels with provenance. It does not decide whether the work was worth it for you.
- Not a billing page. A provider total is a sum. Assay keeps phases, attribution, and a trail for every number.
The three records it joins
| Record | What it is |
|---|---|
| Span | One unit of metered work: tokens by phase, cost, who directed it. |
| Turn | One line of the conversation, sealed at rest. |
| Sitting | One harness session. This is the spine everything else hangs on. |
Two ways people use it
- Your own work with AI — capture the coding harness you already run, then read spend and conversations in the console. Labelling is part of this path.
- Observability in an agent loop — a product records each model call as it happens. Labelling is optional. The debug read is
assay turns.
Next
- Install — invite, binary,
assay init. - After install — capture, console, sweep, labels.
- How it works — the data flow.
- Agents: paste https://docs.assay.solutions/llms.txt into the agent. Human page: For agents.