Pomerado (MIT) turns a website, including one without an API, into an MCP integration your agent can call. It marks each integration read or write. That says whether a tool can change something, not whether this change is fine. Claros adds the second part. The guard sits in front of Pomerado’s MCP servers like any other, checks each write call against your team’s rules and answers stop, ask or nudge before the call runs. Reads pass straight through.

​
What gets checked

The guard reads Pomerado’s own readOnlyHint on each tool, so nothing here is Pomerado-specific. See Read vs write.

​
Setup

  1. Add Pomerado to your client as its README says. Keep OPENAI_API_KEY where Pomerado’s client settings put it: the guard leaves env, envFile and env_vars alone and passes its environment through to Pomerado.
  2. Put the guard in front of it:
  3. After each mint, add the integration’s server from its README.md, then run init again with its name, for example npx getclaros --servers acme_refund. Each integration is its own MCP server, so each one is wrapped on its own.
npx getclaros undo puts every config back as it was.

​
Limits

  • A write mint performs the action once in Pomerado’s own browser while it builds. Claros sees the mint call and can stop it or ask a person before it starts, but it doesn’t see the browser steps inside the mint. Pomerado’s Guardian reviews those steps.
  • After the mint, every call to the write integration goes through the guard like any other write.