Most code-search servers only read. Here the write path is the design: findings, design rationale and research go into a scratchpad that later sessions — and sub-agents, which start with no history at all — can search semantically. The README ships the CLAUDE.md text to make an assistant use it that way, including the three-step habit of searching first, falling back to plain file tools second, and storing what was learned at the end. Note the rebuild in progress before you commit a large index to it.
A Qdrant-backed search layer with automatic Git project detection, a Rust daemon doing the file watching and embedding, and a Rust CLI for operations. Search combines semantic similarity with keyword matching using Reciprocal Rank Fusion; code is chunked with Tree-sitter and enriched with LSP data for active projects.
- Hybrid semantic and keyword search across indexed content, plus exact substring and regex search over FTS5 — `search`, `grep`
- Direct document lookup by id or metadata filter, and file and folder listing for exploring project structure — `retrieve`, `list`
- A write path: store content, register projects, and save notes so findings survive the session — `store`
- Persistent behavioural rules that load back into context at the start of a session — `rules`
- Four collections with different scopes: indexed project code, a library of external reference docs, a scratchpad for analysis and design rationale, and the rules themselves
- A code relationship graph with PageRank, community detection and betweenness centrality, queried from the CLI
- A background daemon that watches files continuously, so the index tracks the working tree without manual re-ingestion
A running Qdrant reachable at `QDRANT_URL` — the README's one-liner starts the `qdrant/qdrant` image on port 6333 with a storage volume. A C compiler, because Tree-sitter grammars ship as C source and are compiled locally — Xcode Command Line Tools on macOS, build-essential on Debian or Ubuntu, Visual Studio Build Tools on Windows. Clang and LLVM only if you build the daemon from source; prebuilt binaries do not need it. Install through Homebrew, the download script, or a source build; that puts `wqm`, `memexd` and the server on your PATH. `QDRANT_API_KEY` is needed for Qdrant Cloud, and `FASTEMBED_MODEL` selects the embedding model. Apache 2.0 licensed, and a v0.2.0 rebuild is in progress.
One command — workspace-qdrant-mcp