Context Lens
A browser extension that reads the page you're on and explains posts and threads in context — multi-provider AI, visible pipeline, tested.
- Type
- Browser extension
- Year
- 2026
- Stack
- TypeScript · Vite · Vitest
Understanding a post or a thread on X, Reddit, or YouTube means copy-pasting into a chat window and losing all the context. The information that makes a post make sense — who is replying to whom, what the parent said — never makes the trip.
- 01
Every platform structures its DOM differently, and redesigns break selectors.
- 02
AI calls cost money and fail — the extension must show its work and degrade honestly.
- 03
API keys belong to the user, not the extension.
- 01
Platform adapters (X / Reddit / YouTube / generic fallback) parse each site into one canonical node model — the rest of the system never knows which site it's on.
- 02
A context engine builds a graph and resolves task-specific depth under a token budget: caption tasks pay for minimal context, understanding tasks get the full package.
- 03
A model router with automatic provider fallback (Gemini → Groq → NVIDIA → OpenRouter), BYOK — keys stay local, no accounts, no telemetry.
- 04
A visible pipeline: every run shows reading page → found target → built context → asking provider → done/failed with reason. Results render as observed / inference / speculation.
Canonical model over per-site logic
Each platform is parsed once into {id, type, text, author, parentId, children} nodes. Tasks and the AI layer consume only the canonical graph — adding a platform means writing one adapter, not rewriting features.
Cost scales with the task
The context engine packages different context for different tasks, so a lightweight caption request doesn't pay for deep-thread context it never uses.
- ✓
Vitest test suite covering the adapters, the context engine, and the AI layer.
- ✓
TypeScript throughout, typechecked build (tsc --noEmit) as part of the workflow.