Coroid can connect to MCP servers you run, giving agents tools beyond reading and writing code — querying an internal service, looking something up in a system of record, calling an API only you have.
Configure under Settings → Connections → External MCP, which has Servers, Templates and Analytics.
What this is for
The gap it closes is context Coroid cannot read from your repository. An agent that can query your service catalogue, check a feature flag's real state, or look up which team owns a module can make decisions it would otherwise have to guess at.
Adding a server
Register the server with its endpoint and credentials. Coroid reads its tool schemas so agents know what is available and how to call it.
Templates cover common server configurations so you are not starting from scratch.
Automation policies
Policies govern what may be called and under what circumstances — particularly what is allowed to happen without a human approving it.
Set these before enabling a server in anger. The question to answer explicitly: which of these tools would I be uncomfortable with an agent calling at 3am without asking? Those need approval; the rest can run unattended.
Analytics
Shows which tools are actually being called and how often.
Worth checking after adding a server. Tools that are never called are not earning their configuration; a tool called far more than expected is worth understanding before it becomes load on the system behind it.
Scoping
Prefer narrow tools over general ones. "Look up the owner of a service" is safer and more useful than "run an arbitrary query", because the narrow version cannot be misused and its results are predictable enough for an agent to reason about.