Agents rebuild what they cannot see. saer mcp, a local MCP server, hands them the system Saer reconstructed from your code: before a task, saer_recommend says which existing components to build it from, with the evidence; after, saer_check reports drift on the lines the agent wrote. It is deterministic (no model at answer time) and read-only. It is in the beta.
A real exchange, on the founder's app
The task, as an agent would receive it: "add a confirm-delete dialog for removing a saved birth chart". saer_recommend answered: build it from Modal for the dialog and Button with variant="destructive" for delete, and treated BirthChart as context, the thing the task acts on, not the UI to build.
Then saer_check on a draft that hard-coded the brand purple, an off-scale radius and a button too small to tap with its outline removed: it named the token to use instead, the nearest radius steps, the spacing class, the missing focus style and the undersized target. The full exchange, verbatim, is in the case study.
Captured from saer mcp on Lunary on 22 October 2026. In the beta.
Why it is trustworthy
- Deterministic: the same repository and task give the same answer. No model at answer time.
- Evidence first: each recommendation says what it matched on and where the component lives.
- Local and read-only: nothing leaves your machine, and the server changes nothing.
What Saer does about this
system.json(in the beta): a machine-readable record of the colour system Saer found, which you can give an agent now.- Scanning after an agent writes code shows any colour drift it introduced (in the beta).
saer mcp: components, hooks, tokens, recommendations and change checks for Claude Code, Cursor and Codex.
- A published benchmark of agent output with and without Saer: in development; results are published with their full method once reviewed.
Every capability and its state: what Saer checks.
See it on your own codebase.
Request early access