docsGround

Why docsGround?

AI models hallucinate when coding with modern or rapidly evolving libraries. Traditional docs MCP servers are slow, require bloated Python environments or heavy headless browsers, and consume massive cloud resources.

docsGround solves this by delivering an ultra-fast, zero-cloud, fully self-hosted documentation grounding stack running on Bun + SQLite FTS5 + Local ONNX BGE Embeddings.

• 1.4s Deep Crawl: HTTP/2 AST crawler indexes hundreds of pages in seconds (no Playwright overhead).

• Hybrid Retrieval: BM25 exact symbol matching combined with 384-dimensional dense semantic vectors.

• 100% Offline & Private: Built-in quantized BGE-Small ONNX vectorizer running locally on CPU.

• Live Web Hybrid Fallback: Auto-queries DuckDuckGo & Brave for queries outside the local index.

• Notion-Style UI: Modern dark dashboard with per-library Markdown readers and live background progress cards.

• Autonomous Agent MCP Suite: 8 native MCP tools allowing AI agents to self-serve, index, and manage docs.

Related MCP server: tech-doc-mcp

Quick Start

Requires Bun 1.3+.

git clone https://github.com/youssefvdel/docsGround.git
cd docsGround
bun install

# Start the Web UI + REST API (default port 3030)
bun run src/index.ts serve

# Run as a Stdio MCP server (default mode)
bun run src/index.ts mcp

# Ingest a library from the CLI
bun run src/index.ts ingest tauri https://docs.rs/tauri/latest/tauri/

Open http://localhost:3030 (or your LAN IP http://192.168.x.x:3030) in your browser.

Embedding model

The default embedding provider is the built-in local ONNX BGE-Small model — fully offline, zero API cost, ~15ms per document. You can switch to a remote OpenAI-compatible gateway from the UI (Settings → Embedding Provider), e.g. http://127.0.0.1:20128/v1.

Autonomous MCP Tool Suite

Register with an MCP client

{
  "mcpServers": {
    "docsground": {
      "command": "bun",
      "args": ["run", "/absolute/path/to/docsGround/src/index.ts", "mcp"]
    }
  }
}

Architecture

flowchart LR
    classDef input fill:#a5d8ff,stroke:#1971c2,color:#0b2545
    classDef core fill:#b2f2bb,stroke:#2f9e44,color:#0b3d1e
    classDef store fill:#c3fae8,stroke:#0ca678,color:#063d2d

    UI["React Web UI :3030"]:::input
    MCP["MCP Server (stdio)"]:::input
    CLI["CLI ingest"]:::input

    API["Bun.serve REST API"]:::core
    API --> E["Search Engine"]:::core
    API --> J["Job Manager"]:::core
    API --> C["Config Manager"]:::core

    E --> DB[("SQLite FTS5 + Vector BLOB")]:::store
    E --> SE["SearxNG Multi-Engine Meta Search"]:::core
    E --> EM["Local ONNX BGE Embeddings"]:::core

    F["StealthFetcher (HTTP/2 + LinkeDOM)"]:::core
    F --> P["DocParser"]:::core
    P --> DB

    UI --> API
    MCP --> API
    CLI --> API

REST API Reference

• GET /api/search?q=<query>&library=<lib>&limit=10 — Hybrid semantic + FTS5 search with live-web fallback.

• GET /api/libraries — List all indexed libraries and page counts.

• GET /api/library-docs?library=<lib> — List all document paths for a library.

• GET /api/doc?id=<doc_id> — Fetch one indexed document's full content.

• POST /api/ingest — Trigger background non-blocking documentation indexing with live progress tracking.

• GET /api/jobs — Poll active and completed ingestion tasks with progress percentage.

• GET /api/config / POST /api/config — View and update runtime configuration.

• POST /api/web-search — Live multi-engine meta-search.

• POST /api/extract — Fetch any URL and convert to clean Markdown.

Development

bun test        # run the test suite
bun run src/index.ts serve   # local dev server

Configuration lives in ~/.docsground/config.json (created automatically on first run).

License

MIT © 2026 docsGround Contributors