Your agent can search and read your files. Lens returns what a search cannot: where the change belongs, which existing piece to extend, which pattern to skip, which library you already have. Connect once. Every session after that, the agent gets that decision before it writes a line — fewer correction loops, less back-and-forth, less senior-engineer review time, and lower AI consumption from repeated discovery and retries.
Lens turns your codebase into memory. Decisions come from that — not a fresh pile of files sent to the model every time.
Task: "add a refund flow to billing."
Three separate misses from one keyword search — a dead file that matches, a live consumer that doesn't, and a pattern worth not repeating. Memory catches all three at once.
Lens comes out of building Helix's own indexing pipeline, MCP integration layer, and control plane from scratch. We didn't set out to make AI write more code faster. We set out to stop re-explaining the same system boundaries in every pull request review.
That's the whole premise: the fix isn't a smarter model. It's giving it the context an engineer would give a new hire on day one.