Pick the measurement that fits the problem.

Spend drift

What your agent costs

Wrap your Anthropic, OpenAI or Gemini client with one line per route. Cost attributes every euro to a route, feature or agent. Then it replays your real traffic against a cheaper model, has an independent judge score the answers across five dimensions, and recommends the switch only once the cheaper model matches on 95% of them.

• One line per route. TypeScript and Python. Wrap the client you already use.
• Every euro attributed. Real cost in euros, mapped to a route, feature or agent.
• Verified against your own traffic. You apply the change. Cost does not switch anything for you.

Quality drift

Whether it is still right

Eval will turn a system prompt into a test suite you can edit, then score real production output against it every day and tell you which check started failing, with the examples that failed. None of it is built yet.

• A prompt will become a test suite. Editable checks, not a score out of ten.
• Will be scored daily against real output. A pass rate, and the production examples that failed it.
• Will not need the SDK. It will read output your agent already saves, in a repository, a database or a document store.

In development. Nothing to sign up for yet

Not published yet. Nothing to visit.

Exposure drift

Whether it is safe

Give Watch an agent URL and it grades the surface you expose, A to F, against the OWASP Agentic Top 10 and NIST IR 8596. It goes deeper under signed authorisation, and re-checks as your prompts, tools and models change.

• A grade, and the report behind it. A to F, with the findings that produced it.
• Mapped to the frameworks. Every finding tied to the OWASP Agentic Top 10 and NIST IR 8596.
• Deeper under signed authorisation. No scanning or exploitation without written permission, scoped in advance.