A self-hosted AI agent built entirely inside n8n

What it does

n8n-claw is a self-hosted autonomous agent constructed entirely from n8n workflows. It converses over Telegram or a generic webhook API, manages tasks and reminders, and maintains long-term memory through a hybrid RAG system that blends semantic search, full-text search, and entity matching with time decay. Under the hood, it also automates browsers, generates new MCP skills on demand, and delegates work to specialized sub-agents—all orchestrated visually in n8n with PostgreSQL as the data layer.

The interesting bit

The entire architecture—heartbeat schedulers, memory consolidation, knowledge graph traversal, even the MCP builder—is implemented as interconnected n8n nodes rather than a custom codebase. That turns the visual workflow editor into something closer to an operating system for agentic behavior, where new capabilities are added by drawing connections rather than deploying services.

Key highlights

• Hybrid memory with RRF fusion, automatic entity extraction, and graceful degradation to keyword search when embedding APIs fail.
• Browser automation with pooled Chromium sessions and interactive 2FA: the agent pauses at a login prompt, asks you for the TOTP code via Telegram, and resumes the same live page.
• Self-expanding MCP skill system that can install catalog skills or build new API integrations from scratch on demand.
• Proactive background workflows (heartbeat every 5 minutes, reminder runner every minute, daily memory consolidation) that keep the agent moving without human prompts.
• Bundled web search via SearXNG and web reading via Crawl4AI, requiring no additional API keys.

Caveats

• The Telegram bot is open to anyone who finds it until you manually add your Chat ID to the allowed list in the n8n UI.
• Browser session persistence is strictly in-memory, so pooled logins evaporate on restart.
• Despite a single setup script, you still need a Linux VPS, a domain or sslip.io workaround, and several API keys to get running.

Verdict

A solid fit for n8n power users who want agentic memory and delegation without abandoning the visual editor. Less appealing if you prefer lightweight, code-first frameworks or want something that runs entirely in-memory without a VPS.