n8nClaw

A lightweight, self-hosted AI assistant built entirely in n8n. Inspired by OpenClaw — multi-channel messaging, persistent memory, task management, and autonomous work — all in a single visual workflow.

Architecture

                        TRIGGERS
          +-----------+-----------+-----------+
          | Telegram  | WhatsApp  |   Gmail   | Hourly
          | Trigger   | Webhook   |  Trigger  | Heartbeat
          +-----------+-----------+-----------+
                |           |           |           |
          [ Filter ]  [ Filter ]  [ Get User ] [ Get User ]
                |           |           |           |
          [ Get User Profile from Init Table ]
                |           |           |           |
          [ Edit Fields — normalize: user_message, system_prompt, last_channel ]
                \           |           |           /
                 +----------+-----------+----------+
                            |
                      +-----v------+
                      |  n8nClaw   |  (Claude Sonnet 4.5 via OpenRouter)
                      |  AI Agent  |  (Postgres Chat Memory - 15 msgs)
                      +-----+------+
                            |
               +------------+-------------+
               |            |             |
          [ Switch: last_channel ]
               |            |
          Telegram     WhatsApp
          Reply        Reply (Evolution API)

                      TOOLS & SUB-AGENTS
          +-------------------------------------------+
          | Tasks DB    | Subtasks DB  | Init/User DB  |
          | Research    | Email Mgr    | Doc Manager   |
          | Worker 1    | Worker 2     | Worker 3      |
          | Vector Store (Supabase RAG)               |
          +-------------------------------------------+

                    MEMORY PIPELINE (scheduled)
          +-------------------------------------------+
          | Postgres Chat History                      |
          |   -> Aggregate messages                    |
          |   -> Summarize (Haiku 4.5)                 |
          |   -> Embed (OpenAI)                        |
          |   -> Store in Supabase Vector DB           |
          +-------------------------------------------+

What's Included

Channels (Triggers)

Core Agent — n8nClaw

• Model: Claude Sonnet 4.5 (via OpenRouter)
• Memory: Postgres chat history (15-message context window)
• Personality: Configurable via "soul" field (name, vibe, purpose)
• User Profile: Living document updated as the agent learns about you

Media Handling (Telegram)

Tools

Sub-Agents

Long-Term Memory Pipeline

A scheduled workflow that:

1. Pulls new chat history from Postgres
2. Aggregates and summarizes conversations (Haiku 4.5)
3. Embeds summaries (OpenAI embeddings)
4. Stores in Supabase vector database for RAG retrieval

Setup

Prerequisites

• n8n (self-hosted or cloud)
• Accounts/API keys for the services you want to use

1. Import the Workflow

1. Open n8n
2. Go to Workflows > Import from File
3. Select n8nClaw.json

2. Create Data Tables

Create three n8n data tables:

Init Table (user profile):

Tasks Table:

Subtasks Table:

3. Configure Credentials

Set up the following credentials in n8n (only configure what you need):

4. Update Placeholders

Search the workflow JSON for YOUR_ and replace with your actual values:

5. Set Up Supabase Vector Store

Create a documents table in Supabase with the pgvector extension enabled. The table should match the schema expected by n8n's Supabase Vector Store node (with a match_documents function).

6. Activate

1. Connect all credentials in the n8n UI
2. Update the data table IDs to point to your tables
3. Update the filter nodes with your Telegram chat ID / WhatsApp number
4. Activate the workflow

Customization

Adding a New Channel

1. Add a new trigger node (webhook, poll, etc.)
2. Add a filter node to validate incoming messages
3. Add a "Get row(s)" node to fetch user profile from the Init table
4. Add an "Edit Fields" node to normalize into user_message, system_prompt_details, and last_channel
5. Connect to the n8nClaw agent
6. Add a new output in the Switch node to route responses back to the channel

Changing Models

All models are configured via OpenRouter. To swap models, edit the model ID in any OpenRouter Chat Model node. The tiered worker system is:

• Worker 1: Cheap/fast model for simple tasks
• Worker 2: Mid-tier model for moderate complexity
• Worker 3: Best model for reasoning-heavy work

License

MIT