OpenCodeIntel

OpenCodeIntel is a code-search platform for AI coding agents: a web app, a REST API and a 12-tool MCP server that give an agent real context on a codebase instead of letting it guess.

Indexing parses each repo with tree-sitter into function-level chunks (Python, JavaScript and TypeScript), embeds them and stores them in Pinecone next to a BM25 index. On the API and MCP search path, a query runs both, fuses the results with reciprocal rank fusion and reranks the top results (Voyage rerank-2 in the research eval, Cohere rerank-v3.5 in production). The public playground runs a lighter path: vector search with keyword and code-graph boosts, no model reranking. The MCP server speaks stdio for local agents and streamable HTTP for hosted Claude.ai connectors.

On a 665-query research eval across 14 open-source codebases, retrieval hit 94% Hit@1, and a 98-run ablation isolated the reranker at +8.4 points. On the public playground, a cold search takes 641ms at the median and a cached repeat 242ms. One result went against me: rerankers trained on web text made code search worse. It stays in the research log.