We’ve been working closely with teams building AI systems for the software engineering stack - autocomplete bots, PR review agents, and code assistants. One common pain point we see: these systems hallucinate about dependencies.
While most companies index the primary codebase with grep, semantic search, and AST-based tools, dependencies are often overlooked. Even though they typically make up the bulk of running code. Maintaining source ingestion into agents is often difficult: fetching tarballs, caching, and avoiding latency spikes.
So we built Package Search MCP. It ingests and indexes public dependencies from NPM, PyPI, Go, and Crates.io at distinct versions. We:
1. Use Tree-sitter to parse and chunk the code
2. Embed the data with Qwen3-Embedding
3. Index each version into its own fork on Chroma Cloud
4. Expose grep + semantic search tools for agents via MCP
Just add this MCP server to your coding agent or AI SDK (Cursor, Claude, Codex, OpenAI etc.), and your agent immediately gets smarter about dependencies. Prompt it with “Use package search” and it’ll know where to look.
We’re excited to see what people build with it - try it out and let me know what you think!
