The personalization layer for the next generation of AI apps
The protocol to give AI apps your life's context.

Whether it's an agent like Poke or a chat app like Claude, next gen AI apps rely on accurate personal context. This raises an issue at both sides, for the AI app and the user.

For AI apps:

• Engineering and maintaining a system to gather and use personal context eats up engineering time, ends up working okay at best in practice, and pulls focus away from shipping the core product.

For users:

• Linking all of your connectors (email, slack, iMessage, etc) to every AI app you use is high-friction, a privacy risk, and leaves your own context fragmented and inaccessible.

This gap is only getting bigger as we move to an AI-adjusted world.

LocalBrain bridges this gap; it automatically organizes personal context from all your connectors into a local, readable knowledge base that any AI app can query to safely understand you.

Search Query:

User types "conferences attended"
  ↓
Frontend POST /protocol/search {"q": "conferences attended"}
  ↓
Daemon receives query
  ↓
Agentic search: LLM generates grep pattern "conference|attended|event"
  ↓
Execute ripgrep on vault files
  ↓
Read relevant file sections
  ↓
Synthesize answer with citations
  ↓
Return JSON response with results + metadata

Ingestion:

Connector fetches new data (e.g., Gmail emails)
  ↓
Convert to ConnectorData format (title, content, timestamp, source_url)
  ↓
LLM analyzes: "Where does this belong in the vault?"
  ↓
Generate structured markdown with ## sections
  ↓
Fuzzy match existing files/sections (tolerance for typos)
  ↓
Apply changes to vault files
  ↓
Validate markdown structure (title, citations, sections)
  ↓
If errors: regenerate with feedback (max 3 retries)
  ↓
Save citation metadata to .json sidecar

LocalBrain is a three-layer system: Electron frontend (macOS app) → FastAPI daemon → hybrid markdown vault. Also an optional MCP proxy server enables AI apps to query the vault.

1. FastAPI Daemon

• Main service running as background process
• Handles agentic search, ingestion, and connector management
• Auto-syncs connected data sources every 10 minutes
• Stateless HTTP API with CORS for frontend access

2. Agentic Search Engine

• Uses Claude Haiku (claude-haiku-4-5-20251001) with tool calling
  • we chose this model since its fast, cheap, and accurate, but it can be swapped out for any LLM
• Tools: grep_vault (ripgrep-based regex search) and read_file
• LLM decides search strategy: decompose query → generate patterns → grep files → read relevant sections → synthesize answer
• No vector embeddings, no similarity scoring—pure regex + LLM reasoning
• 95% accuracy on LongMemEval benchmark (19/20 questions)
  • this is a random sample of questions from the benchmark, not a full evaluation
• We took inspiration from how SoTA coding agents retrieve the most relevant info while being blazingly fast

3. Ingestion Pipeline

• LLM analyzes raw data (emails, messages, docs) and updates the structured markdown filesystem to include the new info if its releavant to the user
• Fuzzy matching for section/file names using Levenshtein distance
• Validation feedback loop: attempts ingestion → checks markdown structure → retries if errors (max 3 attempts)
• Citations tracked in .json sidecars with source URLs, timestamps, and metadata

4. Connector Plugin System

• We made a standardized connector framework, so all connector plugins work nicely and are relatively easy to develop
• Source can either be external (over the web, like Gmail, Discord, etc) or pull from a local source (browser history, iMessage database, etc)
• Auto-discovery: drop <name>_connector.py in connectors/<name>/ and it's loaded on startup
• Interface: BaseConnector with 4 methods (get_metadata, has_updates, fetch_updates, get_status)
• Generic REST routes (/api/connectors/<id>/sync, /status, etc.) work for all connectors

5. MCP Proxy Server

• This is how AI apps can safely query your local filesystem knowledge base
• Pure format translator—zero business logic
• Bridges Claude Desktop (stdio) ↔ Daemon (HTTP)
• Handles authentication (API keys) and audit logging
• Tools exposed to Claude: search, open, summarize, list
• Packaged as .mcpb extension for one-click Claude Desktop installation

6. Electron Frontend

• Next.js app wrapped in Electron for native desktop experience
• Real-time status indicators for daemon and MCP server health
• Resizable panels: file tree, editor, chat, connections, notes
• Dark mode with shadcn/ui components and Tailwind CSS

No vector database for search:

• Ripgrep is instant (<100ms on 10K files)
• LLM generates optimal search patterns (better than embedding similarity)
• Zero indexing overhead, works on any markdown vault
• Transparent: see exactly what matched via grep results

LLM-powered ingestion:

• Handles ambiguity and context (e.g., "Q3 launch" → finds correct project section)
• Self-correcting via validation loops (95%+ success rate)
• Maintains human-readable markdown structure
• No brittle rules or templates—adapts to any content

Plugin architecture:

• Add new connectors without touching daemon code
• Generic API routes scale to infinite connectors
• Easy testing: each connector is isolated

MCP as pure proxy:

• All intelligence in daemon (single source of truth)
• MCP just translates formats (no duplicate logic)
• Easy to debug: test daemon directly, MCP is transparent layer

Markdown as storage:

• Human-readable and editable
• Git-friendly (version control, diffs, branches)
• Portable (works with any markdown editor)
• No vendor lock-in, no database corruption

• Search latency: 1-3s (ripgrep ~50ms + LLM calls ~200ms each)
• Ingestion speed: ~5s per item (LLM analysis + fuzzy matching + validation)
• Memory footprint: ~200MB (FastAPI + Anthropic SDK)
• Disk usage: Vault size + ~10% overhead for citation JSON files
• Concurrent requests: FastAPI handles 100+ RPS easily

Backend:

• FastAPI (async Python web framework)
• Anthropic SDK (Claude Haiku API client)
• ripgrep (Rust-based regex search, 100x faster than grep)
• Levenshtein (fuzzy string matching for section names)
• python-dotenv (environment configuration)

Frontend:

• Next.js 15 (React SSR framework)
• Electron 33 (native desktop wrapper)
• TailwindCSS (utility-first styling)
• shadcn/ui (component library)
• Motion/Framer Motion (animations)

Integration:

• Model Context Protocol (Claude Desktop stdio bridge)
• OAuth 2.0 (Gmail authentication)
• Discord.py (Discord API wrapper)

localbrain/
├── electron/
│   ├── app/                        # Next.js frontend
│   │   ├── src/
│   │   │   ├── app/page.tsx       # Main app layout
│   │   │   └── components/        # React components
│   │   └── package.json           # Frontend deps
│   │
│   └── backend/                    # Python backend
│       ├── src/
│       │   ├── daemon.py           # Main FastAPI service
│       │   ├── agentic_search.py   # Search engine (LLM + ripgrep)
│       │   ├── agentic_ingest.py   # Ingestion pipeline (LLM + fuzzy match)
│       │   ├── connectors/         # Plugin system
│       │   │   ├── base_connector.py
│       │   │   ├── connector_manager.py
│       │   │   ├── gmail/
│       │   │   ├── browser/
│       │   │   └── ...
│       │   ├── core/
│       │   │   ├── mcp/            # MCP proxy server
│       │   │   │   ├── server.py
│       │   │   │   ├── stdio_server.py
│       │   │   │   └── tools.py
│       │   │   └── ingestion/      # Ingestion utilities
│       │   └── utils/              # Shared utilities
│       └── requirements.txt
│
└── my-vault/                       # Markdown knowledge base
    ├── projects/
    ├── personal/
    └── ...

Agentic Search Prompt Strategy:

• Ultra-concise system prompt (OpenCode-inspired)
• Example-driven: shows LLM exactly how to use tools
• "Minimize output, answer directly" → reduces token usage
• Forces LLM to check line numbers before reading full files

Fuzzy Matching Algorithm:

• Levenshtein distance with configurable threshold (default: 0.7 similarity)
• Tries exact match first, falls back to fuzzy if no match
• Prevents duplicate sections from slight name variations

Validation Loop:

• After ingestion: parse markdown, check for required sections (# title, ## Related)
• Verify citation markers [1] match entries in .json file
• If errors found: pass to LLM with specific error messages
• Max 3 retries → fail gracefully with detailed error log

Connector Auto-Discovery:

• Scan connectors/ directory for *_connector.py files
• Import and instantiate classes inheriting from BaseConnector
• Register REST routes dynamically using FastAPI's router system
• Maintain singleton ConnectorManager for lifecycle management

MCP Extension Packaging:

• stdio_server.py copied into extension/server/ directory
• manifest.json declares tool schemas (JSON Schema format)
• package.sh creates .mcpb bundle (ZIP with manifest)
• Claude Desktop loads bundle, spawns stdio server subprocess

This architecture optimizes for transparency (see what's happening), simplicity (minimal abstractions), and extensibility (easy to add connectors/features). The markdown vault is the single source of truth, everything else is stateless logic.

Made with ❤️ by Henry Wang, Sid Songirkar, Taymur Faruqui, and Pranav Balaji