AIBRIEF
MCP is the standard that lets AI agents call a web app's functions; Frigade's Hacker News Show HN post proposes building one by watching the app talk to its own backend.
Most web apps still don't expose agent hooks, and hand-writing an MCP-style server for each one doesn't scale. A Show HN from Frigade bets the fix is to watch the app talk to its own backend and auto-generate the integration: recipes, auth, and all.
MCP, the Model Context Protocol, is the emerging standard that lets AI agents call a web app's functions the way a developer would. Most apps have no MCP server at all, and the ones that do are maintained by hand. Frigade's pitch is to skip both: a browser-side agent runs inside the user's authenticated session, observes the live API calls an app already makes on its own behalf, and turns them into reusable tools for an LLM.
Each endpoint becomes a "recipe": method, auth (with token and cookie refresh), input and response schema, and a description. When the host app changes its API, the agent is designed to figure it out and rewrite the recipe, per Frigade.
The friction is real. The agent only works inside a session the user is already logged into, and the app owner still has to approve each discovered tool from a Frigade dashboard before the agent can act. Live demos for Jira, Spotify, Hacker News, and a full walkthrough are linked from the post; the closing example is truncated, so end-to-end UX, pricing, and availability are not yet visible.