Week 5: The layer lives where the code lives
Corexi moves deeper into the IDE. MCP marketplace submissions, one-click installs, and why being inside the agent changes everything.
The most important architectural decision in Corexi wasn't about the engine. It was about where the engine delivers its output.
The insight
A UX tool in a browser tab is a context switch. The developer stops writing code, opens another app, reads findings, goes back to the IDE, and manually translates recommendations into changes. Every step in that chain is a dropout point.
A UX layer inside the AI agent is zero context switches. The agent asks Corexi mid-conversation: "review this checkout form." Corexi returns structured findings with fix-ready code. The agent applies the fix. The developer approves or rejects. No tab. No copy-paste. No dropout.
What we shipped
MCP server with 9 tools. Not a single "run audit" endpoint — a full toolkit. get_ux_rules before writing code. review_code before committing. trigger_scan after deploying. get_findings when the PM asks "what's the UX status."
Six IDEs, one-click install. Cursor, VS Code, Claude Desktop get deeplinks (cursor://, vscode:mcp/install). Claude Code, Windsurf, Replit get copy-paste JSON snippets. The user's real API key is embedded in the install link — from "new key" to "agent sees Corexi" is genuinely one click.
Marketplace submissions. Cursor Marketplace and cursor.directory, both submitted. Being discoverable where developers already browse for tools matters more than any landing page optimization.
Public playground. /mcp/examples lets anyone paste a snippet and watch review_code produce findings live. No login, no key, rate-limited to keep costs manageable. Seeing is believing — this is the fastest way for a skeptical developer to understand what Corexi does.
Why this matters for defensibility
Every tool call through MCP is a data point. Which tools get called most. Which IDEs dominate. Which findings get applied. This telemetry feeds the engine's Phase 2 prediction capabilities — and it only accumulates for tools that are inside the loop.
A browser-tab audit tool doesn't know if its findings were applied. Corexi does.
Enjoyed this? Get the weekly build log in your inbox.