Vlt-Bridge

Work in Progress

Agent memory and vault bridge in one repo: the vlt CLI (threads, CodeRAG, Oracle) and the Document-MCP web app (Obsidian-style vault + MCP server). They share the vault concept — markdown on disk — and that’s it. No runtime dependency: agents shell out to vlt or talk to the MCP server over HTTP; the CLI does not route through the backend.

This is a RAG for any document. A memory system for agents. And a web chat provider to give agents a phone a friend feature.

Audience: Developers and MCP integrators wiring agents to memory/vault; CLI users; anyone running the web UI.

Quick start by goal

Goal	Action
Use the web UI	Backend: cd backend && uv run uvicorn src.api.main:app --reload --host 0.0.0.0 --port 8000. Frontend: cd frontend && npm install && npm run dev. UI at http://localhost:5173, API at :8000.
Use vlt CLI (threads, CodeRAG, Oracle)	cd packages/vlt-cli && pip install -e ".[oracle]", then vlt profile init and vlt config set-key <OPENROUTER_KEY>.
Connect an MCP client	Prefer vlt-mcp (threads + code + oracle + vault in one server): claude mcp add --scope user vlt vlt-mcp. For vault-only: Document-MCP STDIO/HTTP (see MCP Client Configuration).

Architecture

flowchart LR
  subgraph CLI ["vlt CLI"]
    threads[threads]
    coderag[CodeRAG]
    oracle[Oracle]
  end
  subgraph Backend ["Backend"]
    FastAPI[FastAPI]
    MCP[MCP vault tools]
  end
  subgraph Frontend ["Frontend"]
    React[React UI]
  end
  Vault[("Vault (markdown on disk)")]
  CLI --> Vault
  Backend --> Vault
  Frontend --> FastAPI
  FastAPI --> MCP

Loading

CLI and backend both read/write the vault independently. Oracle can run locally (CLI) or in thin-client mode against the backend when it’s running.

Monorepo layout

Vlt-Bridge/
├── backend/              # FastAPI (Document-MCP) + FastMCP STDIO/HTTP
│   └── src/
│       ├── api/          # REST: notes, search, graph, RAG, TTS, auth, oracle
│       ├── mcp/          # MCP server (vault tools only)
│       ├── models/       # Pydantic schemas
│       ├── prompts/      # Oracle prompts
│       └── services/    # vault, indexer, auth, database, rlm_oracle, repl_executor,
│                        # project_context, ans, openrouter_client, ...
├── frontend/             # React 19 + Vite 7 + shadcn/ui
├── packages/
│   └── vlt-cli/          # vlt CLI + vlt-mcp (threads, coderag, oracle, vault)
├── specs/                # SpecKit feature specs
├── data/                 # Runtime data (vaults/, index.db) — gitignored
└── desktop-app/          # Optional Tauri desktop wrapper

vlt CLI

Installed from packages/vlt-cli; primary interface for agent memory and Oracle. Requirements: Python 3.11+, OpenRouter API key, universal-ctags for Oracle (apt install universal-ctags).

Install:

cd packages/vlt-cli
python -m venv .venv && source .venv/bin/activate
pip install -e ".[oracle]"
vlt profile init
vlt config set-key <YOUR_OPENROUTER_KEY>

Thread memory

Threads: per-project, append-only logs. Librarian daemon compresses nodes into semantic state (LLM). thread push is tuned for <50ms so agents can log without blocking.

vlt overview
vlt thread new <project> <thread-id> "goal"
vlt thread push <thread-id> "insight or decision"
vlt thread read <thread-id>
vlt thread read <thread-id> --search "jwt"
vlt thread seek "concept or question"
vlt tag <node-id> "#bug"
vlt link <node-id> <thread-id>

Use --author "AgentName" on push for multi-agent attribution.

CodeRAG

Hybrid indexing: vector + BM25 + ctags call-graph. Languages: Python, TypeScript, TSX, JavaScript, Go, Rust.

vlt coderag init --project <id> --path /path/to/repo
vlt coderag init --project <id> --foreground
vlt coderag init --project <id> --force
vlt coderag status --project <id>
vlt coderag search "authentication flow" --project <id>
vlt coderag map --project <id>
vlt coderag map --project <id> --scope src/api/

Status: pending, running, completed, failed, cancelled. Indexing uses the background daemon by default; if the daemon isn’t running, you’re prompted for foreground. Configure with coderag.toml in project root (e.g. include/exclude).

Oracle

Multi-source query over code index, vault notes, and thread history. RLM Oracle: REPL-centric harness. The LLM gets a Python sandbox where the project is exposed as project, sub_oracle, and Final; it writes code to explore and synthesize answers. Loop ends when it sets Final. Sandbox: RestrictedPython, 30s timeout, approved stdlib only. No Behavior Tree or XML signals.

vlt oracle "How does authentication work?"
vlt oracle "Where is UserService defined?" --source code
vlt oracle "Why did we choose SQLite?" --source threads
vlt oracle "Explain the architecture" --local
vlt oracle "Hard question" --thinking --model anthropic/claude-sonnet-4

Default: Oracle tries the backend (thin-client, shared with web UI). --local forces local. --explain shows retrieval traces. --source can be code, vault, or threads (repeatable). See specs/022-rlm-oracle/quickstart.md for backend-side testing.

Daemon management

vlt daemon start
vlt daemon stop
vlt daemon status
vlt librarian start

vlt-mcp (unified MCP server)

One server for threads, code, oracle, and vault. No separate daemon/ports for core use. Setup: claude mcp add --scope user vlt vlt-mcp. Cold-start ~164ms. Full setup (Claude Desktop, per-project override, oracle toggle): specs/018-vlt-mcp-server/quickstart.md.

Document-MCP

Web UI

Browser vault viewer/editor; talks to FastAPI on port 8000.

• Wikilinks [[Note Name]] via SQLite slug matching (same folder preferred, then lexicographic).
• Full-text search: FTS5, BM25, title 3× weight, recency bonus.
• Backlinks and graph view (force-directed).
• Split-pane editor with live preview; AI chat (RAG + Gemini); TTS (ElevenLabs).
• Optimistic concurrency: if_version on save, 409 on conflict.
• Auth: HF OAuth in space mode, or local with LOCAL_USER_ID.

MCP server

Seven tools, STDIO or HTTP+JWT: list_notes, read_note, write_note, delete_note, search_notes, get_backlinks, get_tags. Vault only; for threads/CodeRAG/Oracle use the CLI or vlt-mcp.

Running locally

Use the working manual commands (start-dev.sh may need adjustment for your env):

# Backend
cd backend
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"
uv run uvicorn src.api.main:app --reload --host 0.0.0.0 --port 8000

# MCP STDIO (Claude Desktop/Code)
uv run python src/mcp/server.py

# MCP HTTP (JWT)
uv run python src/mcp/server.py --http --port 8001

# Frontend
cd frontend
npm install
npm run dev   # http://localhost:5173

Tests: Backend: cd backend && uv run pytest (or tests/unit, tests/integration, -k test_vault_write). Frontend: cd frontend && npm run lint (and npm run build if you want a production check).

MCP client configuration

vlt-mcp (recommended)

cd packages/vlt-cli && pip install -e ".[oracle]"
claude mcp add --scope user vlt vlt-mcp

Document-MCP backend (vault-only, legacy)

STDIO:

{
  "mcpServers": {
    "document-mcp": {
      "command": "uv",
      "args": ["run", "python", "src/mcp/server.py"],
      "cwd": "/absolute/path/to/Vlt-Bridge/backend"
    }
  }
}

HTTP (HF Space + JWT):

{
  "mcpServers": {
    "document-mcp": {
      "transport": "http",
      "url": "https://YOUR_USERNAME-Document-MCP.hf.space/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_JWT_TOKEN"
      }
    }
  }
}

JWT: from the web UI Settings after HF OAuth login.

Environment variables

Backend and local dev. Source of truth: backend/src/services/config.py and backend/.env.example.

Variable	Required	Default	Description
JWT_SECRET_KEY	yes	—	Min 16 chars. python -c "import secrets; print(secrets.token_urlsafe(32))"
VAULT_BASE_PATH	no	./data/vaults	Per-user vault root
ENABLE_LOCAL_MODE	no	true	When true, single-user local mode
LOCAL_USER_ID	no	local-dev	User in local mode
HF_OAUTH_CLIENT_ID	space	—	HF OAuth client ID
HF_OAUTH_CLIENT_SECRET	space	—	HF OAuth client secret
GOOGLE_API_KEY	optional	—	Gemini (RAG / AI chat)
ELEVENLABS_API_KEY	optional	—	TTS
ELEVENLABS_VOICE_ID	optional	—	ElevenLabs voice
ORACLE_MAX_TURNS	no	30	Max agent turns per Oracle query
ORACLE_PROMPT_BUDGET	no	8000	Token budget for Oracle system prompt
BASE_URL	no	http://localhost:8000	For widget/assets in production

SQLite index path defaults to ./data/index.db (not overridden by env in current code). For full list and security notes (e.g. ENABLE_NOAUTH_MCP), see backend .env.example.

Deployment

Backend + frontend are packaged as one Docker image for Hugging Face Spaces. See DEPLOYMENT.md for HF Space creation, OAuth, secrets, and push.

docker build -t vlt-bridge .
docker run -p 7860:7860 -e JWT_SECRET_KEY="dev-secret" vlt-bridge

Extras

• ANS: Backend has an Agent Notification System (tool/budget alerts into the Oracle UI). See CLAUDE.md.
• ChatGPT widget: Served at /widget.html for embedding in ChatGPT; MCP remains available for other agents.
• desktop-app: Optional Tauri app in desktop-app/.
• Specs: Feature specs live under specs/ (SpecKit). Deep architecture and implementation notes: CLAUDE.md.