vibe-hnindex

vibe-hnindex

Local MCP server — index your repo once, search it in every AI session

Keyword (SQLite FTS5) · Semantic (Qdrant + Ollama embeddings) · Hybrid — your code stays on disk

MCP server (vibe-hnindex) latest: v0.12.0 · hnindex-cli v0.12.0 — Docs · Changelog · GitHub Releases

What this does

vibe-hnindex is a Model Context Protocol server. After you index a folder once, assistants (Claude, Cursor, Windsurf, Antigravity, …) can search that codebase with paths and line ranges — data is stored locally (SQLite + optional Qdrant). Embeddings use Ollama; vectors use Qdrant (Docker, local, or Qdrant Cloud with QDRANT_API_KEY).

Documentation

📚 Full docs site: docs.hnindex.cloud — 16 pages covering Getting Started, Configuration, Tools Reference, Guides, and Code Agent.

Page	What you'll learn
Introduction	What vibe-hnindex does, key features, how it works
Installation	Node, Ollama, Qdrant setup + MCP config
Quick Start	5-minute walkthrough with CLI + agent skill
Configuration	All 25+ env vars with embedding model comparison
Search	6 search modes, regex, fuzzy, streaming, cache
Code Agent 🆕	code_session + code_apply with safety scopes
Setup MCP	Per-platform config (Claude, Cursor, Antigravity, VS Code...)

Also available in-repo: docs/getting-started.md, docs/configuration.md, docs/tools-reference.md.

CLI installer (hnindex)

Optional — writes the MCP JSON for you (merge-safe, same npx -y vibe-hnindex block as in the docs):

npm install -g hnindex-cli

# Setup MCP config
hnindex init --mcp antigravity    # or: claude, cursor, windsurf, vscode, codex
hnindex init --list               # show all targets and paths

# Install AI agent skill (recommended)
hnindex init-skill --target claude    # or: antigravity, cursor, windsurf, vscode
hnindex init-skill --list             # show all skill targets

# Update
hnindex update                    # npm update -g hnindex-cli

See docs.hnindex.cloud for full documentation.

Install in 5 steps

• Node.js — v20+ (nodejs.org). On Windows, Node 20 or 22 LTS is strongly recommended so npm install does not need a C++ compiler. See Troubleshooting → Windows if npm i vibe-hnindex fails.
• Ollama — install from ollama.com, then: ollama pull bge-m3:567m and keep ollama serve running (or set OLLAMA_URL to a remote server).
• Qdrant — for semantic/hybrid search: docker run -d --name qdrant -p 6333:6333 qdrant/qdrant (or use Qdrant Cloud). Keyword-only search works without Qdrant.
• MCP config — add the server to your assistant’s MCP settings. Minimal example (self-hosted Qdrant):

{
  "mcpServers": {
    "vibe-hnindex": {
      "command": "npx",
      "args": ["-y", "vibe-hnindex"],
      "env": {
        "OLLAMA_URL": "http://localhost:11434",
        "OLLAMA_MODEL": "bge-m3:567m",
        "QDRANT_URL": "http://localhost:6333",
        "SEARCH_STREAM_ENABLED": "true",
        "CODE_AGENT_ENABLED": "true",
        "CODE_AGENT_SCOPE": "moderate",
        "CHAT_MEMORY_ENABLED": "true"
      }
    }
  }
}

• Restart the IDE or assistant, then in chat ask to index a path and search — see First steps.

For Qdrant Cloud, add QDRANT_API_KEY and set QDRANT_URL to your HTTPS cluster URL — details in Getting started.

Optional rerank (RERANK_URL)

Semantic/hybrid search already uses Ollama (OLLAMA_URL, OLLAMA_MODEL e.g. bge-m3:567m) for query vectors and Qdrant for retrieval. After that, the server can reorder the top pool of hits:

• Without RERANK_URL: reorder by Qdrant semantic scores (no extra network service). This is enough for most setups, including when you only run Ollama + Qdrant.
• With RERANK_URL: POST JSON { "query", "documents" } to your URL; response { "scores": number[] } (same length as documents). Use a small HTTP service you host that wraps your reranker; Ollama does not expose this contract on :11434 by default.

Ollama vs rerank: pulling a reranker model in Ollama (e.g. qllama/bge-reranker-v2-m3) does not replace RERANK_URL—you still need an adapter service unless you only rely on the built-in Qdrant reorder. See Configuration → Rerank.

Env	Role
SEARCH_RERANK	false disables post-retrieval reorder entirely (default: enabled).
SEARCH_RERANK_POOL	Max candidates considered before trim (default 50).
RERANK_URL	Full URL of your {query, documents} → {scores} API (optional).
RERANK_TIMEOUT_MS	Timeout for that POST (default 15000).

Timeouts

To prevent hanging when Ollama or Qdrant are unresponsive, vibe-hnindex applies timeouts on all external calls. You can tune these via environment variables:

Env	Default	Controls
OLLAMA_TIMEOUT_MS	30000 (30s)	Max wait for Ollama /api/embed and /api/tags calls
QDRANT_TIMEOUT_MS	15000 (15s)	Max wait for Qdrant API calls (search, upsert, etc.)
SEARCH_TIMEOUT_MS	60000 (60s)	Overall timeout for the entire search operation

Set any of these to a higher value if you have a slow machine or large dataset. Set to 0 to disable the timeout for that layer (not recommended).

Google Antigravity

Use the same mcpServers block as above, but save it in Antigravity’s MCP file:

File	mcp_config.json under .gemini/antigravity/ in your user folder
Windows	C:\Users\<your-username>\.gemini\antigravity\mcp_config.json
macOS / Linux	~/.gemini/antigravity/mcp_config.json
UI	⋮ menu → MCP → Manage MCP Servers → View raw config

Step-by-step: Integrations → Google Antigravity.

Features (short)

Search	6 modes: keyword (FTS5+BM25), semantic (Qdrant vectors), hybrid (RRF fusion), regex, symbol, auto
Code Agent	code_session — 1 call replaces 5-15 searches. code_apply — safe code changes with auto test/lint/typecheck
Chat Memory 🆕	Auto-track tool calls, semantic search via Qdrant, persistent AI context across sessions
Streaming	Parallel keyword+semantic search (~1.5-2× faster), 4-phase progress notifications
Fuzzy Search	Levenshtein distance auto-corrects typos ("fucntion" → "function")
Smart Context	Task-aware context: impact analysis, test file detection, similar code patterns
Storage	SQLite on disk + Qdrant for vectors; 100% local, no cloud required
Indexing	Incremental (SHA-1 hash), parallel workers (~3-4× faster), watch mode (auto re-index on save), 40+ languages, .hnindexignore
Resilience	Keyword search works without Qdrant or Ollama; graceful degradation
Benchmark	Built-in benchmark_search tool — compare streaming vs non-streaming, all search modes
Multiple Embedding Models	bge-m3 (default), nomic-embed-text, qwen3-embedding, mxbai-embed-large, and more

Architecture

graph TB
    subgraph Input["📂 Input"]
        A["💻 Your Codebase<br/>.ts .py .go .rs ..."]
    end

    subgraph Server["⚙️ vibe-hnindex MCP Server"]
        B["🔍 Search Router<br/>keyword | semantic | hybrid"]
        C["🔀 RRF Fusion"]
    end

    subgraph Storage["💾 Storage"]
        D[("SQLite<br/>FTS5 + Keyword")]
        E[("Qdrant<br/>Vector Embeddings")]
    end

    subgraph Memory["🧠 Chat Memory (v0.12)"]
        F[("SQLite<br/>Chat Context")]
        G[("Qdrant<br/>Chat Vectors")]
    end

    subgraph Infra["🏗️ Infrastructure"]
        H["Ollama<br/>Embeddings"]
        I["Qdrant<br/>localhost:6333"]
    end

    subgraph Output["🤖 AI Clients"]
        J["Claude · Cursor · Windsurf<br/>Antigravity · VS Code"]
    end

    A -->|"index_codebase"| Storage
    A -->|scan| H
    B -->|"keyword"| D
    B -->|"semantic"| E
    B -->|"hybrid"| C
    C --> D
    C --> E
    B -.->|"auto-track"| F
    F --> H
    H --> G
    D --> J
    E --> J
    H -.-> I

    style F fill:#6366f1,color:#fff
    style G fill:#6366f1,color:#fff
    style B fill:#f59e0b,color:#000
    style J fill:#22c55e,color:#fff

How indexing & search work →

License

MIT — see LICENSE.

Contributing

Issues and PRs: github.com/AndyAnh174/vibe-hnindex.

Contact

Ho Viet Anh (AndyAnh174) · [e-mail] · GitHub