ClawLink — Cross-Instance Agent Communication

ClawLink turns isolated OpenClaw sessions into a collaborative agent mesh. Multiple OpenClaw instances — on the same machine or across a network — can delegate tasks, share findings, co-edit files, and coordinate work.

Security note: The relay server only runs when you explicitly start it with python3 server.py. It stores no data on disk — everything is in-memory and clears on shutdown. When connecting across machines, use --token to require authentication.

Architecture

  Machine A (OpenClaw)          Machine B (OpenClaw)
       │                              │
       └──── HTTP/WebSocket ──────────┘
                    │
            ┌───────┴───────┐
            │  ClawLink      │
            │  Relay Server  │
            │  (localhost or │
            │   LAN machine) │
            └────────────────┘

One machine runs the relay server. All others connect as agent clients. The relay handles message routing, queuing, and agent registry — entirely in-memory.

Quick Start

Step 1: Start the Relay Server

On the machine that will act as the hub (can be one of the agent machines):

# Install dependencies (user install — no system changes)
pip install --user aiohttp requests

# Start the relay — localhost only (safest for single-machine use)
python3 scripts/server.py

# Or bind to LAN with auth token for multi-machine use
python3 scripts/server.py --host 0.0.0.0 --token YOUR_SHARED_SECRET

The server defaults to 127.0.0.1 (localhost only). To allow other machines on your local network, add --host 0.0.0.0 --token YOUR_SECRET.

For access from another machine on your network:

Use the relay machine's local IP address directly — e.g., http://192.168.1.10:9077. Both machines must be on the same network and you must pass the same --token on both server and clients.

Step 2: Register This Agent

python3 /path/to/clawlink/scripts/client.py \
  --relay http://RELAY_IP:9077 \
  --token YOUR_SHARED_SECRET \
  register \
  --name "DESCRIPTIVE_NAME" \
  --caps "COMMA_SEPARATED_CAPABILITIES" \
  --description "What this agent specializes in"

The token can also be set via CLAWLINK_TOKEN environment variable.

Step 3: Discover and Communicate

python3 scripts/client.py --relay http://RELAY_IP:9077 --token SECRET discover

Core Operations

Discovering Peers

python3 scripts/client.py discover

Returns a table of online agents with their IDs, names, capabilities, and machines.

Delegating Tasks

python3 scripts/client.py delegate \
  --to TARGET_AGENT_ID \
  --task "Clear description of what needs to be done" \
  --context '{"key": "relevant context data"}' \
  --priority normal

Receiving and Responding to Tasks

# Poll for incoming messages
python3 scripts/client.py poll

# Respond when done
python3 scripts/client.py respond \
  --to REQUESTING_AGENT_ID \
  --msg-id ORIGINAL_MESSAGE_ID \
  --result "Task result or summary of work done"

Broadcasting Knowledge

python3 scripts/client.py broadcast \
  --content "Description of the finding" \
  --topic "category" \
  --tags "tag1,tag2"

Collaborative File Editing

# Upload/update a shared file
python3 scripts/client.py file-put --key "report.md" --file ./report.md

# Download a shared file
python3 scripts/client.py file-get --key "report.md" --output ./report.md

# List shared files
python3 scripts/client.py file-list

Behavioral Guidelines for Agents

As a Task Receiver

1. Poll regularly — Check for messages every 30-60 seconds during active work, or when the user asks "any messages?" or "check ClawLink"
2. Acknowledge receipt — Respond with status "in_progress" when you get a task
3. Be thorough — Complete the full task before responding
4. Report failures — If you can't complete a task, respond with status "failed" and explain why

As a Task Delegator

1. Match capabilities — Use discover to find the right agent for the job
2. Provide context — Include file paths, URLs, constraints, and output format
3. Be patient — The other agent may take time

General

• Always tell the user what's happening on the network
• Surface incoming messages proactively
• Keep heartbeats alive during long sessions

Troubleshooting

Protocol Reference

For the full message format specification, transport layer details, and workflow patterns, read references/protocol.md.