Glimind // glossary

Plain definitions for AI-agent reliability — what these terms mean and why they matter.

MCP server — An MCP (Model Context Protocol) server exposes tools, resources, and prompts that an AI agent can discover and call over a standard protocol.
tool reliability — Tool reliability is whether an AI agent tool (an MCP tool or API) is working correctly right now — measured from real-usage success rate, latency, and error patterns, not just whether it is reachable.
MCP gateway — An MCP gateway (or tool router) sits between agents and many MCP servers, brokering discovery and tool calls with auth, policy, and sometimes reliability-aware failover.
pre-flight check — A pre-flight check is a quick query an agent makes before calling a tool to learn whether it is healthy, how to shape the input, and what to use if it is down.
working recipe — A working recipe is a currently-successful, value-free input shape for a tool, distilled from real successful calls — so an agent knows how to call the tool correctly right now.
reliability score — A reliability score is a neutral 0–100 rating (with an A–F grade) summarizing how reliably an AI agent tool is working, from cross-ecosystem real usage and safe liveness probes.
schema drift — Schema drift is when a tool's advertised input/output schema changes, which can silently break agents that relied on the old shape.
A2A protocol — A2A (Agent2Agent) is a protocol for AI agents from different providers to discover and message each other as peers, advertising capabilities via an agent card.