Obi★ FinalistDeep-reviewed
by himavanth karpurapu
What they're building
Local-first context engine for developers. Indexes your code and files once, then hands your AI assistant only the chunks it needs instead of re-scanning the whole repo. Hybrid on-device retrieval, MCP handoff to Cursor and Claude Desktop. Built with Nebius, Composio, and Tavily for BuilderShip.
AI code reviewrepo: real
Real, substantial local-first Electron/React app: hybrid retrieval (FTS5 lexical + sqlite-vec semantic + CLIP image embeddings), OCR routing, local llama-server sidecars (gemma/nomic gguf), a local tool-calling loop, and a working stdio MCP server exposing obi_search to Cursor/Claude Desktop. The core retrieval engineering is genuinely implemented. However, the build_description and README headline 'Built with Nebius, Composio, and Tavily,' yet a grep across src/electron/mcp-server/tools/ipc finds those names ONLY in README.md — none are imported or called anywhere — so sponsor_usage is floored. Inference is 100% on-device; the cloud/sponsor tier is aspirational. ⚑ Sponsor stack is named-only — zero Tavily/Nebius/Composio in code; package.json has none of them. Demo is a YouTube video (page un-scrapeable) and the app is local-only Electron, so the head-to-head claim is not independently verifiable. Submitted repo_url path differs from the actual clonable repo (Personal_Vault/Local-First-Multimodal-File-Index-via-MCP).