The claim the project leads with is token cost: returning matched chunks instead of grep hits plus whole files. Indexing an average repo takes about 500 ms and queries about 1 ms on CPU, with NDCG@10 of 0.854 on their own benchmarks — self-reported, but measurable on your own repo since there is nothing to sign up for.
A code search library built for agents, exposed as an MCP server. Ask it a question in plain language and it returns the matching chunks rather than whole files, which is what makes it cheap enough for an agent to use constantly. Everything runs locally on CPU: no API keys, no GPU, no external service.
- A codebase searched with a natural-language or code query, against a local path or an `https://` git URL, scoped to `code`, `docs`, `config` or `all` — `search`
- Chunks semantically similar to a given file and line, for finding related implementations — `find_related`
- On-demand indexing with caching, and automatic re-indexing of local paths when files change
- A standalone CLI for scripts and non-MCP use, including searching a remote repo that is cloned on demand, and a `--max-snippet-lines` control for how much of each hit comes back
- Index control through `.gitignore` and `.sembleignore`, where a `!` prefix force-includes extensions it would otherwise skip
- An interactive installer that detects your agents and offers three integrations: the MCP server, CLI guidance written into AGENTS.md or CLAUDE.md, and a dedicated `semble-search` sub-agent
Uv, then `uv tool install semble` and `semble install` to wire it into your agents; `semble uninstall` reverses it. No API keys, no GPU, no external services — everything runs on CPU. Indexes are built and cached on first run and invalidated automatically when files change. MIT licensed.
One command — uv tool install semble semble install