Resources
An MCP server is enterprise-ready when an agent can call its tools without gaining implicit standing, leaking secrets, or executing attacker-controlled side effects. This rubric is a scoring sheet, not a marketing claim about any specific server.
Model Context Protocol servers sit on the boundary between language models and real systems: files, tickets, cloud APIs, and sometimes shells. Grade each server before you connect it to production agents, then enforce those controls at runtime with UnitOne Gateway.
Every tool call must be attributable to a service identity or user. Reject anonymous transports in production. Prefer short-lived credentials, SSO, or workload identity over long-lived static tokens embedded in agent configs.
Expose the smallest tool surface the workflow needs. Separate read tools from write/execute tools. Bind authorization to the calling principal and tenant, not to 'the agent is internal.' Dangerous tools (shell, arbitrary HTTP, unrestricted file write) should be off by default.
Secrets must not appear in tool descriptions, sample prompts, traces, or model context. Responses should redact credentials, session tokens, and personal data. If a tool fetches customer content, define retention and whether that content can re-enter a prompt.
Treat tool arguments, retrieved documents, and third-party web content as untrusted. Do not let retrieved text choose which tool runs next without policy. Validate schemas, canonicalize paths, and block unexpected protocols or destinations.
Pin egress to known destinations. Disable unused transports. If the server can run code or commands, sandbox it, drop privileges, and cap CPU, memory, and time. Record the execution host so incidents are investigable.
Log authentication outcomes, tool name, caller, policy decision, and a redacted argument digest. Alert on bursty tool use, denied calls, and first-time use of high-risk tools. Keep a kill switch that revokes a server or tool without redeploying every agent.
Pin versions, review dependency diffs, and require a reviewed change for new tools or permission expansions. Third-party servers need a named owner, an update channel, and a tested rollback. Intent-preserving fixes should land as pull requests against the server repo, not silent hotpatches.
Start free to inspect and block unsafe MCP tool use at runtime. See plans for Team and Enterprise event volume. A standalone MCP scanner is legacy and is not on the pricing grid.