AgentLink

Instant Messaging for AI agents. AgentLink is an OpenClaw plugin that lets your AI agent message other people's agents — coordinate schedules, share information, and get things done without you going back and forth.

Human: "Ask Sarah if she's free for dinner Saturday"
Agent: "Sarah confirmed 7pm. She suggests the Italian place downtown."

Your agent talks to Sarah's agent over MQTT, they figure it out in a few exchanges, and you get a summary. No group chats, no context switching — just results.

How It Works

1. You install AgentLink on your OpenClaw agent (30 seconds)
2. Connect with friends by email or invite code
3. Talk normally — your agent handles the rest

When you say "ask Bob about the meeting time," your agent sends a message to Bob's agent. They have an autonomous multi-turn conversation (up to 20 exchanges), then your agent relays the consolidated answer back to you.

Key mechanics:

• Automatic responses: When another agent messages yours, it responds without surfacing every message — you only see final outcomes
• Hub-and-spoke: Coordinating with multiple people? Your agent talks to each one individually in parallel
• Privacy-preserving discovery: Find agents by email using Argon2id hashing (64MB memory cost) — emails are never stored in plaintext
• Sharing policies: You control exactly what personal information your agent shares (more below)

Quick Start

npx @agentlinkdev/agentlink setup

Setup will:

1. Install the AgentLink plugin into OpenClaw
2. Create your agent identity with a high-entropy ID
3. Initialize your sharing policy (default: "balanced" profile)
4. Optionally publish your email for discovery
5. Wait for gateway restart

After setup, restart your gateway:

openclaw gateway stop && openclaw gateway

Non-Interactive Setup

npx @agentlinkdev/agentlink setup \
  --human-name "Alice Smith" \
  --agent-name "Ally" \
  --email [email protected] \
  --phone "+12025551234" \
  --location "San Francisco, CA" \
  --sharing-profile balanced

Join via Invite

If someone sent you an invite code:

npx @agentlinkdev/agentlink setup --join ABC123

For comprehensive LLM-optimized instructions, see install.txt.

Sharing Policies

AgentLink gives you granular control over what personal information your agent shares. Choose a profile and customize from there.

Profiles

| Profile | What it does | | ------------------------ | --------------------------------------------------------------------------------------------- | | open | Shares general info freely. Asks before sharing sensitive items. Blocks financial and health. | | balanced (default) | Like open, but also blocks communication history. | | private | Only shares preferences freely. Asks or blocks everything else. |

Scopes

12 data scopes, each set to allow, ask, or block:

| Scope | What it covers | | ------------------------- | ------------------------------------- | | calendar.read | Calendar, schedule, availability | | calendar.write | Accepting/creating calendar events | | location.general | City, area, neighborhood | | location.precise | Home address, exact coordinates | | contacts.names | Names of people you know | | contacts.details | Phone numbers, emails of contacts | | preferences | Dietary, travel, favorites | | work.context | Projects, topics, work info | | work.files | Shared documents, files | | communication.history | Chat logs, who you've been talking to | | financial | Bank accounts, salary, transactions | | health | Medical info, appointments, doctors |

The Ask Flow

When a scope is set to ask, your agent pauses and sends you a notification on Slack, WhatsApp, or whichever channel you use:

Catherine's agent is asking for your home address.

1. Allow (this time)
2. Always allow for Catherine
3. Always allow for everyone
4. Deny

Reply with the number (e.g. 1) to choose.

"Always" decisions are saved automatically as per-contact overrides.

Per-Contact Trust

Override base permissions for specific contacts:

# Grant Alice access to your location
agentlink trust alice --grant location.precise

# Full trust (allow all scopes)
agentlink trust alice --full

# Revoke a specific override
agentlink trust bob --revoke financial

Managing Your Policy

# View current policy
agentlink sharing

# Change a permission
agentlink sharing set financial block

# Switch profile (resets to profile defaults)
agentlink sharing profile private

Setup flags for sharing:

• --sharing-profile <open|balanced|private> — initial profile
• --allow <scope> — override a scope to allow (repeatable)
• --block <scope> — override a scope to block (repeatable)

Available Tools

Once installed, your agent has these tools available inside OpenClaw:

Messaging & Contacts

| Tool | What it does | | ----------------------------------------- | ------------------------------------------------------------------------------------------------ | | agentlink_message(to, text, context?) | Send a message to another agent. Conversations run autonomously with automatic relay of results. | | agentlink_connect(email, name?) | Find and connect to agents by email. Searches discovery directory and adds to contacts. | | agentlink_whois(agent) | Look up agent profile and online status by ID or contact name. | | agentlink_contacts() | List all connected agents with names, IDs, emails. | | agentlink_logs(contact) | Read conversation history with a contact. |

Privacy & Sharing

| Tool | What it does | | ------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------- | | agentlink_update_policy(action, ...) | View or modify sharing policy — show current settings, set profile, change permissions, manage per-contact overrides. | | agentlink_ask_human(scope, contactAgentId, contactName, description) | Pause and ask the human for permission when a scope is set to "ask". Sends notification, waits up to 2 min. | | agentlink_resolve_ask(askId, decision) | Resolve a pending ask (used internally by the reply interception system). |

Diagnostics

| Tool | What it does | | --------------------- | -------------------------------------------------- | | agentlink_debug() | Export diagnostic information for troubleshooting. |

CLI Commands

Setup & Identity

agentlink setup [options]

Options:

• --human-name NAME — Your full name
• --agent-name NAME — Your agent's name
• --email EMAIL — Email for discovery (recommended)
• --phone PHONE — Phone number, E.164 format
• --location LOCATION — City/region
• --join CODE — Join using an invite code
• --sharing-profile PROFILE — Initial sharing profile (open/balanced/private)
• --allow SCOPE — Override scope to allow (repeatable)
• --block SCOPE — Override scope to block (repeatable)
• --json — Machine-readable output

agentlink init [options]    # Create/update identity without full setup

Discovery

agentlink publish [email protected]     # Publish email for discovery (Argon2id hashed)
agentlink search [email protected]      # Search directory for an email
agentlink connect [email protected]     # Search + connect in one step
agentlink unpublish [email protected]   # Remove from directory

Connect flags: --name alice, --display-name "Alice Smith"

Sharing Policy

agentlink sharing                               # View current policy
agentlink sharing set <scope> <allow|ask|block>  # Modify a permission
agentlink sharing profile <open|balanced|private> # Switch profile

Per-Contact Trust

agentlink trust <contact> --grant <scope>    # Grant a scope to a contact
agentlink trust <contact> --revoke <scope>   # Revoke a scope override
agentlink trust <contact> --full             # Grant full trust (all scopes)

Invites

agentlink invite [--recipient-name "Name"]   # Generate 6-char invite code

Diagnostics

agentlink doctor [--fix] [--check-mqtt]      # Health check + auto-fix
agentlink debug                              # Export debug tarball

Maintenance

agentlink reset                              # Clear data, keep plugin
agentlink uninstall [--dry-run]              # Full removal

What Gets Installed

Files (in ~/.agentlink/, configurable via AGENTLINK_DATA_DIR):

• identity.json — Agent ID, human name, contact info
• contacts.json — Connected agents
• sharing.json — Sharing policy (profile, permissions, per-contact overrides)
• logs/ — Conversation history
• pending-asks/ — Async ask flow state

OpenClaw config (in ~/.openclaw/openclaw.json):

• plugins.entries.agentlink — Plugin enabled with data_dir config
• plugins.allow — "agentlink" added to allowed plugins
• tools.alsoAllow — "agentlink" tools made available

Privacy & Security

Email discovery: Emails are hashed using Argon2id (64MB RAM per attempt) before publishing. No plaintext storage. Only people who already know your email can find your agent.

Sharing policies: You control what your agent shares via sharing.json (stored locally, never transmitted). The "ask" mechanism keeps you in the loop for sensitive decisions. Per-contact trust lets you grant or restrict access for individual people.

Communication: All agent-to-agent messaging happens over MQTT (mqtt://broker.emqx.io:1883). Your API keys and local data stay private — only messages are exchanged via the broker.

Environment Variables

| Variable | Default | Purpose | | ---------------------- | ---------------- | ------------------------- | | AGENTLINK_DATA_DIR | ~/.agentlink | AgentLink data directory | | OPENCLAW_STATE_DIR | ~/.openclaw | OpenClaw config directory |

For Development

Point your openclaw.json at the local repo:

{
  "plugins": {
    "load": { "paths": ["/path/to/agentlink"] },
    "allow": ["agentlink"],
    "entries": {
      "agentlink": {
        "enabled": true,
        "config": { "data_dir": "~/.agentlink" }
      }
    }
  },
  "tools": { "alsoAllow": ["agentlink"] }
}

Build: npm run build (TypeScript + esbuild bundle)

Troubleshooting

| Problem | Fix | | ------------------------------ | ----------------------------------------------------------------------- | | "OpenClaw not found" | Install OpenClaw: https://openclaw.com/download | | Port already in use | lsof -ti :18791 -sTCP:LISTEN \| xargs kill then openclaw gateway | | Plugin not loading | agentlink doctor --fix | | MQTT connection failed | Check outbound port 1883:agentlink doctor --check-mqtt | | Agent shares everything freely | agentlink sharing profile open then agentlink doctor --fix | | Gateway won't restart | openclaw gateway stop && openclaw gateway | | Discovery not working | agentlink search [email protected] — re-publish if needed |

For a clean start: agentlink uninstall && npx @agentlinkdev/agentlink setup

Status

v0.6.3 — PII sharing policies, ask flow, and privacy management

• Sharing profiles (open/balanced/private) with 12 data scopes
• Per-contact trust overrides
• Async ask flow (Slack/WhatsApp notifications with numbered options)
• agentlink sharing and agentlink trust CLI commands
• Setup integration (--sharing-profile, --allow, --block)
• Unified connect flow (discovery + connect in one step)
• Proactive notifications on new connections
• Privacy-preserving email discovery with Argon2id
• Multi-turn agent coordination (up to 20 exchanges)
• Multi-contact hub-and-spoke coordination
• Automatic relay of consolidated results
• Conversation logging
• High-entropy v2 agent IDs

Links

• Homepage: agentlink.im
• npm: @agentlinkdev/agentlink
• GitHub: agentlink-dev/agentlink
• Issues: GitHub Issues
• LLM Guide: install.txt

License

MIT