Vibe Hnindex

---

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 (default), OpenAI, Voyage, Gemini, or an OpenAI-compatible API; see Embedding providers. Cloud providers receive the content being embedded; 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 MCP JSON or Codex TOML 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

1. Node.js — v22+ (nodejs.org). CI verifies Node 22 and 24. SQLite uses N-API binaries bundled for supported platforms; see Troubleshooting → Windows if npm i vibe-hnindex fails.
2. Embedding provider — choose OpenAI, Voyage, Gemini or a compatible API, or use local Ollama: install from ollama.com, then: ollama pull bge-m3:567m and keep ollama serve running (or set OLLAMA_URL to a remote server).
3. 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.
4. 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"
      }
    }
  }
}

5. 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.

Agent Workspace (v0.15.0)

workspace_context identifies the active project purpose, Git/index state and a declared task persisted per project/session. locate_code finds files/symbols then keywords locally, returning line evidence and freshness. Explicit hybrid mode uses configured semantic APIs. MCP startup instructions and knowledge://workspace expose this workflow. Project-scoped CLI installs bind their root automatically; global installs follow client roots unless explicitly configured.

See Workspace and code locations. Refresh stale positions before editing; use filesystem tools for unavailable or excluded code.

Code Graph (v0.14.0)

Build a TypeScript/JavaScript symbol graph with index_code_graph(path, project_name) using SQLite alone. index_codebase also builds it; file indexing and watching update relationships automatically. Use find_references, callers, and graph_context to inspect source evidence and bounded relationship context. No graph database service is required.

AST chunking preserves function/class boundaries where possible. After upgrading, run index_codebase once to rebuild chunks and vectors. See Code Graph guide for resolution limits, offline usage, token budgets and quality evaluation.

Optional rerank

Hybrid search combines keyword results and embedding vectors using RRF. Optional reranking supports Voyage (RERANK_PROVIDER=voyage, VOYAGE_API_KEY or RERANK_API_KEY) and custom HTTP (RERANK_PROVIDER=http, RERANK_URL). Without a reranker, or if it fails, the original retrieval ranking is preserved. Exact symbol/regex modes skip reranking.

Env	Role
SEARCH_RERANK	false disables reranking.
SEARCH_RERANK_POOL	Candidate pool before final trim (default 50; at least the requested limit).
RERANK_PROVIDER	none, http, voyage; defaults to http when RERANK_URL is set, otherwise none.
RERANK_MODEL	Voyage model, default rerank-3-lite.
RERANK_URL	Custom endpoint; HTTP contract is {query, documents} -> {scores}.
RERANK_API_KEY	Optional Bearer credential; Voyage can use VOYAGE_API_KEY.
RERANK_TIMEOUT_MS	Full request/body timeout, default 15000.

See Configuration.

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 Providers (v0.13.0)	Ollama, OpenAI, Voyage, Gemini and OpenAI-compatible APIs; provider/model-aware vector collections
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) · hovietanh147@gmail.com · GitHub

Codex and architecture (v0.15.2)

Run npx -y hnindex-cli init --mcp codex inside the project, then optionally npx -y hnindex-cli init-skill --target codex. Codex project MCP uses .codex/config.toml (trusted projects only); skills use .agents/skills/use-vibe-hnindex. Restart/reconnect Codex and ask for workspace_context, followed by index_code_graph.

See CLI configuration, system, sequence and user diagrams, and official Codex MCP setup.