# ADR-033: RuVector + Ruflo MCP Tool Integration **Status:** Accepted **Date:** 2026-03-04 **Context:** chat-ui-mcp MCP Bridge ## Context The MCP bridge initially shipped with 3 built-in tools (search, web_research, system_guide). Users want access to the full ruvector (10 tools) and ruflo (205+ tools) ecosystems from within the HF Chat UI without running separate MCP servers. ### Tool Inventory | Backend | Tools | Categories | |---------|-------|------------| | **ruvector** | 10 | Intelligence (hooks_stats, hooks_route, hooks_remember, hooks_recall, hooks_init, hooks_pretrain, hooks_build_agents, hooks_verify, hooks_doctor, hooks_export) | | **ruflo** | 205+ | Agent (7), Swarm (4), Memory (7), Config (6), Hooks (40+), Task (6), Session (5), Hive-mind (9), Workflow (9), Analyze (4), Progress (4), AIDefence (6), AgentDB (14+) | ## Decision Integrate ruvector and ruflo as **stdio MCP child processes** spawned by the bridge, with tool calls proxied through the existing `/mcp` HTTP endpoint. ### Architecture ``` ┌─────────────────────────────────────────────────┐ │ HF Chat UI (browser) │ │ MCP_SERVERS: http://mcp-bridge:3001/mcp │ └─────────────────┬───────────────────────────────┘ │ JSON-RPC 2.0 over HTTP ▼ ┌─────────────────────────────────────────────────┐ │ MCP Bridge (Express) │ │ │ │ ┌──────────────────┐ ┌─────────────────────┐ │ │ │ Built-in Tools │ │ StdioMcpClient │ │ │ │ • search │ │ ┌───────────────┐ │ │ │ │ • web_research │ │ │ ruvector (10) │ │ │ │ │ • system_guide │ │ └───────────────┘ │ │ │ └──────────────────┘ │ ┌───────────────┐ │ │ │ │ │ ruflo (205+) │ │ │ │ │ └───────────────┘ │ │ │ └─────────────────────┘ │ └─────────────────────────────────────────────────┘ ▲ stdin/stdout (JSON-RPC) ▲ │ │ npx ruvector mcp start npx ruflo mcp start ``` ### Key Design Decisions 1. **Namespaced tool names**: External tools are prefixed with `{backend}__` (e.g., `ruvector__hooks_route`, `ruflo__agent_spawn`) to avoid name collisions with built-in tools. 2. **Lazy startup**: Backends initialize after Express starts listening, so the bridge is immediately available for health checks. If a backend fails to start, built-in tools still work. 3. **Environment toggle**: Each backend can be disabled via `ENABLE_RUVECTOR=false` or `ENABLE_RUFLO=false` for deployments that don't need all tools. 4. **Graceful shutdown**: SIGTERM/SIGINT handlers kill child processes cleanly. 5. **Timeout protection**: Each tool call has a 30s timeout. Backend initialization has a 15s timeout. ## Implementation ### StdioMcpClient A reusable client class that: - Spawns a child process with the MCP server command - Sends JSON-RPC messages over stdin, reads responses from stdout - Manages pending request map with UUID correlation IDs - Handles newline-delimited JSON protocol - Auto-discovers tools via `tools/list` on initialization ### Tool Routing ``` tools/call request → name starts with "{backend}__"? → YES: strip prefix, route to StdioMcpClient.callTool() → NO: route to built-in executeTool() ``` ### Configuration ```env # In docker-compose.yml or .env ENABLE_RUVECTOR=true # default: true ENABLE_RUFLO=true # default: true ``` ## Consequences ### Positive - 215+ tools available from HF Chat UI without separate MCP server management - Single `/mcp` endpoint — no client-side config changes needed - Built-in tools work even if backends fail to start - Namespacing prevents tool name collisions ### Negative - Additional memory/CPU for child processes (~50MB each) - First request may be slow while npx resolves packages - Backend stderr goes to bridge logs (noisy) ### Mitigations - Backends are optional (env toggle) - npx caches packages after first run - Startup is non-blocking ## Related - [ADR-029: HuggingFace Chat UI Cloud Run](ADR-029-HUGGINGFACE-CHAT-UI-CLOUD-RUN.md) - [ADR-030: MCP Tool Gap Analysis](ADR-030-MCP-TOOL-GAP-ANALYSIS.md) - [ADR-032: RVF Private MCP Tunnel](ADR-032-RVF-PRIVATE-MCP-TUNNEL.md)