Claude Context vs grep: setup and code search tests

Claude Context is an MCP server that gives a coding assistant indexed code search. The speaker explains how Tree-sitter splits code into function and class chunks, Merkle DAG hashes track changed files, and semantic search works alongside BM25 keyword matching. The claimed context reduction is up to 40%; the demonstration does not establish that saving for every query.

The setup uses OpenCode with GLM-5 Turbo and Zilliz Cloud, though the speaker says Milvus can run locally for a self-hosted database. This demonstration still uses OpenAI embeddings. It requires Node 20 or newer but below 24, with Node 22 used here. The speaker reports cloud-cluster timeouts and switches to a serverless cluster.

Indexing the roughly 1.5-million-line VS Code repository takes about 50 minutes. A separate 23,000-line repository takes less than a minute to index.

Search results are mixed. The entry-point query takes 19 seconds with Claude Context and 14 seconds with grep and glob. A follow-up requesting exact code takes 23 seconds with the index and 49 seconds without it. For an Electron main-process explanation, the speaker reports about 1 minute 47 seconds versus roughly 5 minutes, with more detail from Claude Context. Sub-agent token usage is not fully visible. The speaker prefers it for repositories of around 20,000 to 30,000 lines, while acknowledging that it is not consistently faster or more token efficient.