Workspace Qdrant MCP
workspace-qdrant-mcp
面向 AI 助手的项目级向量数据库,提供混合语义 + 关键词搜索,并具备自动项目检测功能。
特性
混合搜索 - 使用倒数排名融合(Reciprocal Rank Fusion)结合语义相似度和关键词匹配
项目检测 - 自动识别 Git 仓库并进行项目级集合管理
6 个 MCP 工具 - search, retrieve, rules, store, grep, list
代码智能 - 针对活跃项目的 Tree-sitter 语义分块 + LSP 集成
代码图谱 - 包含算法(PageRank、社区检测、中介中心性)的关系图谱
高性能 CLI - 基于 Rust 的
wqm命令行工具后台守护进程 - 用于持续文件监控和处理的
memexd
Related MCP server: Super-Memory-TS
快速开始
前置要求
Qdrant -
docker run -d -p 6333:6333 -v qdrant_storage:/qdrant/storage qdrant/qdrantC 编译器 - 首次使用时编译 Tree-sitter 语法所需。Tree-sitter 语法以 C 源码形式分发并在本地编译。
macOS:
xcode-select --install(Xcode 命令行工具)Linux:
apt install build-essential(Debian/Ubuntu) 或dnf groupinstall "Development Tools"(Fedora)Windows: 安装 Visual Studio Build Tools 并勾选 C++ 工作负载
安装
选项 1:Homebrew(推荐 — macOS 和 Linux)
brew install ChrisGVE/tap/workspace-qdrant
brew services start workspace-qdrant选项 2:预构建二进制文件
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/ChrisGVE/workspace-qdrant-mcp/main/scripts/download-install.sh | bash
# Windows (PowerShell)
irm https://raw.githubusercontent.com/ChrisGVE/workspace-qdrant-mcp/main/scripts/download-install.ps1 | iex将 wqm 和 memexd 安装到 ~/.local/bin (Linux/macOS) 或 %LOCALAPPDATA%\wqm\bin (Windows)。
选项 3:从源码构建
git clone https://github.com/ChrisGVE/workspace-qdrant-mcp.git
cd workspace-qdrant-mcp
./install.sh详细说明和平台特定注意事项请参阅 安装参考。对于 Windows,请参阅 Windows 安装指南。
配置 MCP
Claude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"workspace-qdrant-mcp": {
"command": "node",
"args": ["/path/to/workspace-qdrant-mcp/src/typescript/mcp-server/dist/index.js"],
"env": {
"QDRANT_URL": "http://localhost:6333"
}
}
}
}Claude Code:
claude mcp add workspace-qdrant-mcp -- node /path/to/workspace-qdrant-mcp/src/typescript/mcp-server/dist/index.js验证
wqm --version
wqm status healthCLAUDE.md 集成
将以下内容添加到您项目的 CLAUDE.md(或全局的 ~/.claude/CLAUDE.md)中,以便 Claude Code 主动使用 workspace-qdrant:
## workspace-qdrant
The `workspace-qdrant` MCP server provides codebase-aware search, a library knowledge base, a scratchpad for accumulated insights, and persistent behavioral rules. The tool schemas are self-describing; these instructions cover *when* and *how* to use them.
### Primary Search and Knowledge Base
**Use `workspace-qdrant` first whenever context is uncertain** — first session on a project, returning after a significant gap, or exploring an unfamiliar subsystem. It is faster and more accurate than walking files manually, and it retrieves findings from prior sessions that would otherwise be lost.
**Three-step protocol:**
1. **Search** with `workspace-qdrant` (`search`, `grep`, `list`, or `retrieve`)
2. **Fall back** to `Grep`, `Glob`, `WebSearch` only when workspace-qdrant is insufficient or unavailable
3. **Store** any new findings, analysis, or design rationale via `store` so they are retrievable in future sessions
When a fresh handover or strong prior context already covers what you need, skip the exploratory search — but always store new findings at the end.
**Collections and their purpose:**
- `projects` — indexed codebase; use `scope="project"` (current project) or `scope="all"` (across all projects)
- `libraries` — external reference docs, API specs, third-party documentation; add via `store` with `collection="libraries"` and search with `includeLibraries=true`
- `scratchpad` — analysis, design rationale, research transcripts, architectural insights; complements session handovers by building a growing, semantically searchable knowledge layer across sessions
- `rules` — persistent behavioral rules; load at session start via `rules` → `action="list"`
**Practical notes:**
- Use `grep` for exact strings or regex; `list` with `format="summary"` to explore project structure
- Store external docs or specs into `libraries` so they are searchable alongside code
- Use the scratchpad to record *why* decisions were made, not just *what* was done — future sessions can retrieve the reasoning
### Sub-Agents
Sub-agents start with only the prompt you give them — they have no session history or handover context. They must always use `workspace-qdrant` first for any code exploration, without exception. Include this verbatim in every agent prompt:
> "You have no prior context about this codebase. Use `workspace-qdrant` as your mandatory first tool for ALL code searches — symbols, functions, architecture, patterns, prior findings. Use `search`, `grep`, `list`, or `retrieve` before touching any file with Read/Grep/Glob. Store any new findings, analysis, or design rationale via `store` (scratchpad for insights, libraries for reference docs) so they persist for future sessions."
### Project Registration
At session start, check whether the current project is registered with workspace-qdrant. If it is not, ask the user whether they want to register it (do not register silently). Once registered, the daemon handles file watching and ingestion automatically — no further action is needed.
### Behavioral Rules
The `rules` tool manages persistent rules that are injected into context across sessions. Rules are **user-initiated only** — add rules when the user explicitly instructs you to, never autonomously. Use `action="list"` at session start to load active rules.
### Issue Reporting
workspace-qdrant is under active development. If you encounter errors, unexpected behavior, or limitations with any workspace-qdrant tool, report them as GitHub issues at https://github.com/ChrisGVE/workspace-qdrant-mcp/issues using the `gh` CLI.MCP 工具
工具 | 用途 |
| 在索引内容中进行混合语义 + 关键词搜索 |
| 通过 ID 或元数据过滤器直接查找文档 |
| 管理持久的行为规则 |
| 存储内容、注册项目、保存笔记 |
| 使用 FTS5 进行精确子字符串或正则表达式搜索 |
| 列出项目文件和文件夹结构 |
参数和示例请参阅 MCP 工具参考。
集合
集合 | 用途 | 隔离方式 |
| 项目代码和文档 | 按 |
| 参考文档(书籍、论文、文档) | 按 |
| 行为规则和偏好设置 | 按 |
| 临时工作存储 | 按会话隔离 |
CLI 参考
# Service management
wqm service start # Start background daemon
wqm service status # Check daemon status
wqm status health # System health check
# Search and content
wqm search "query" # Search collections
wqm ingest file path.py # Ingest a file
wqm rules list # List behavioral rules
# Project and library
wqm project list # List registered projects
wqm project watch pause # Pause file watchers
wqm library list # List libraries
wqm tags list # List tags with counts
# Administration
wqm admin collections list # List collections
wqm admin rebuild all # Rebuild all indexes
wqm admin backup create # Backup snapshots
wqm admin stats overview # Search analytics
# Code graph
wqm graph stats --tenant <t> # Node/edge counts
wqm graph query --node-id <id> --tenant <t> --hops 2 # Related nodes
wqm graph impact --symbol <name> --tenant <t> # Impact analysis
wqm graph pagerank --tenant <t> --top-k 20 # PageRank centrality
# Setup
wqm init completions zsh # Shell completions
wqm init man install # Install man pages
wqm init hooks install # Install Claude Code hooks
# Queue and monitoring
wqm queue stats # Queue statistics完整文档请参阅 CLI 参考。
配置
环境变量
变量 | 默认值 | 描述 |
|
| Qdrant 服务器 URL |
| - | API 密钥(Qdrant Cloud 必需) |
|
| 嵌入模型 |
架构
+-----------------+
| Claude/Client |
+--------+--------+
|
+--------v--------+
| MCP Server | (TypeScript)
+--------+--------+
|
+--------------+--------------+
| |
+--------v--------+ +--------v--------+
| Rust Daemon | | Qdrant |
| (memexd) | | Vector Database |
+--------+--------+ +-----------------+
|
+--------v--------+
| File Watcher |
| Code Graph |
| Embeddings |
+-----------------+Rust 守护进程负责文件监控、嵌入生成、代码图谱提取和队列处理。所有写入操作均通过守护进程路由以确保一致性。
文档
用户指南:
参考:
规格说明、ADR 和开发者资源请参阅 文档索引。
开发
# TypeScript MCP server
cd src/typescript/mcp-server && npm install && npm run build && npm test
# Rust daemon and CLI (from src/rust/)
cargo build --release
cargo test
# Graph benchmarks
cargo bench --package workspace-qdrant-core --bench graph_bench
# Binaries output to:
# - target/release/wqm
# - target/release/memexd贡献
开发设置和指南请参阅 CONTRIBUTING.md。
许可证
MIT 许可证 - 详情请参阅 LICENSE。
灵感来源于 claude-qdrant-mcp
Available Tools
6 toolsgrepB
Search code with exact substring or regex pattern matching. Uses FTS5 trigram index for fast line-level search across indexed files.
Inputs: pattern (required), regex, caseSensitive, pathGlob, +5 more.
listA
List project files and folder structure. Shows only indexed files (excludes gitignored, node_modules, etc). Use format "summary" first to understand project layout, then drill into specific folders with the path parameter.
Inputs: path, depth, format, fileType, +7 more.
retrieveA
Retrieve documents by ID or metadata filter. Use this to access specific documents when you know the document ID. Prefer search for discovery, retrieve for known documents.
Inputs: documentId, collection, filter, limit, +3 more.
rulesA
Manage behavioral rules (add, update, remove, list). Check active rules at the start of each session to load the user's behavioral preferences. Rules persist across sessions and guide how you should work.
Inputs: action (required), content, label, scope, +5 more.
searchA
Search for documents using hybrid semantic and keyword search. Use this tool FIRST when answering questions about the user's codebase, project architecture, or stored knowledge. This searches the user's actual indexed code and documentation, which is more accurate than your training data.
Inputs: query (required), collection, mode, scope, +14 more.
storeA
Store content or register a project. Use type "library" (default) to store reference documentation, type "url" to fetch and ingest a web page, type "scratchpad" to save persistent notes/scratch space, or type "project" to register a project directory for file watching and ingestion.
Inputs: type, content, libraryName, forProject, +8 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 6 tools
Each tool targets a distinct function: grep searches code lines, list navigates files, retrieve fetches known documents, rules manages preferences, search discovers content, and store ingests content. There is no functional overlap.
All tool names are single lowercase verbs (grep, list, retrieve, rules, search, store), following a consistent and predictable pattern.
With 6 tools covering searching, navigation, retrieval, storage, and rule management, the count is well-scoped for a workspace knowledge server without being excessive or sparse.
The tool surface covers the core workflows of searching, browsing, retrieving, storing, and managing rules. One minor gap is the lack of explicit update/delete operations for stored documents, though store may allow overwriting.
Maintenance
Related MCP Connectors
Shared memory for coding agents. Stop re-explaining your codebase every session.
Persistent memory for AI agents. Semantic search, memory graph, W3C DID identity.
Persistent memory and knowledge graphs for AI agents. Hybrid search, context checkpoints, and more.
Universal memory for AI agents and tools. Save, organize and search context anywhere.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenancePersistent knowledge memory layer for AI agents. Hybrid semantic + full-text search with pgvector, code dependency graph with blast-radius impact analysis, and incremental indexing for 7 languages. In-process ONNX embeddings, no external API required.45 PyPI34MIT
- AlicenseNot gradedqualityDmaintenanceLocal-first semantic memory server with project indexing for AI assistants. It enables AI assistants to store, retrieve, and search memories and project code using embeddings and vector search.68 npmMIT
- AlicenseCqualityBmaintenanceGives AI coding assistants persistent project memory and semantic code search, running fully locally with no API keys required.56MIT
- AlicenseNot gradedqualityFmaintenanceIndexes codebases into Qdrant for semantic search, enabling AI assistants to find relevant code by meaning without re-exploring the repo.MIT