NCI indexes TypeScript declarations and package relationships on disk so tools can answer questions from the same sources you compile against. The overview lives in the Introduction; what follows is how this work came about and how it lines up with what we ship today.
A compact mental model helps agents use it well: NCI is an index and query surface over your tree—exports, types, dependency edges—not another runtime. When a tool can ask structured questions instead of rediscovering the same files every turn, the model gets room to think about architecture, trade-offs, and safer ways to implement a change instead of burning context on dead-end greps.
Origin
The story starts with a real Cursor agent session in the editor: local generated types versus web documentation, dead ends, and the moment the workflow lands on something stable enough to ship. The replay below is shortened from that trace so you can follow the beats without sitting through the full clock time.
Why NCI exists
Agents and editors already have access to your source tree, but searching
raw text across node_modules is
brittle. Public docs help with orientation; they do not replace the exact
signatures and overload sets in the version you installed. NCI exists so
those answers can be indexed once and queried quickly, without repeating the
same dead ends on every task.
Who it is for
It is for teams that work in real JavaScript and TypeScript monorepos, and
for anyone wiring an agent through MCP or the CLI who wants stable,
structured access to exports, types, and dependency edges. If you already
maintain skills or prompts that tell an agent to read
.d.ts files by hand, NCI is the same
idea turned into an index you can depend on across sessions.
What it replaces
It replaces open-ended grepping through generated types, guessing which file holds the truth, and falling back to scraped documentation when search misses. Instead of repeating that loop on every question, you index the tree you care about and let the agent pull precise rows from a local database.