Sourcegraph Cody

Cody's pitch is simple: it knows your entire codebase, not just the file you're editing. In practice, that means asking "where does this config value get used?" and getting answers that span 47 repos, not just grep results. The autonomy shows up when you ask it to refactor a method signature - it'll propose changes across call sites in services you forgot existed, then generate a migration plan. I've used it to onboard onto a fintech monorepo where the last engineer left eighteen months ago. Cody surfaced the actual data flow in twenty minutes. A junior would've needed a week. The deep-context engine is the real differentiator. It indexes your code graph, not just embeddings, so it understands dependencies and call hierarchies. When you ask it to add a feature, it knows which modules import what, which tests will break, and where the edge cases live. That's legitimately autonomous in a way Cursor isn't - Cursor guesses from local context, Cody actually knows. Failure modes: it's overkill for small projects. If your repo fits in a single editor window, the indexing overhead buys you nothing. The free tier is hobbled - 20 queries a day, no custom context. The enterprise tier makes sense if you're already paying for Sourcegraph, otherwise the pricing feels steep compared to Cursor or Continue. The UI in VS Code is fine but not delightful - it's clearly built by backend engineers who care more about correctness than polish. Compared to GitHub Copilot Workspace, Cody is less opinionated about workflow but more accurate about existing code. Compared to Cursor, it's slower to start (indexing takes time) but better at multi-file reasoning. I'd reach for Cody when joining a large team or working in a polyglot monorepo. For greenfield projects or solo work, Cursor's speed wins. One concrete workflow: I asked it to trace how user permissions propagate through a Django app. It identified the middleware, the decorator, three database models, and a Redis cache layer, then explained the flow in plain English. That's not autocomplete - that's actual code comprehension.