🤖↔️🤖 Inter-AI Communication: got my NanoClaw agent talking to Claude Code

[davyd26]

What I built

I set up a lightweight communication bridge between my NanoClaw agent and Claude Code — my admin AI that manages the server.

The result: the two AIs can now collaborate asynchronously without human involvement:

• Claude Code can inject messages into the agent's next prompt
• The agent can send requests back to Claude Code (diagnostics, system checks, etc.)
• Claude Code reads pending messages at the start of each session and responds

---

How it works

Two plain log files act as the channel:

File	Direction
groups/main/logs/agent-inbox.log	Claude Code → Agent
groups/main/logs/agent-to-admin.log	Agent → Claude Code

A 12-line patch in src/index.ts checks for inbox content before every agent invocation. If content exists, it's prepended to the prompt and the file is cleared:

// Inject admin inbox: if Claude Code left a message for the agent, prepend it
const inboxPath = path.join(GROUPS_DIR, group.folder, 'logs', 'agent-inbox.log');
try {
  const inbox = fs.readFileSync(inboxPath, 'utf-8').trim();
  if (inbox) {
    prompt = `[Message from Claude Code]\n${inbox}\n[End of Claude Code message]\n\n${prompt}`;
    fs.writeFileSync(inboxPath, '');
    logger.info({ group: group.name }, 'Injected Claude Code inbox message into prompt');
  }
} catch {
  // No inbox file — normal case
}

No polling daemon, no websocket, no extra process. Just filesystem IPC — consistent with how NanoClaw already works.

---

What emerged from this

Once the bridge was in place, the two AIs started collaborating naturally:

• Agent asked Claude Code to check how many containers were running (we had a duplicate container bug causing 3x messages)
• Claude Code monitored RAM usage and set alert thresholds based on info from the agent
• They agreed on a shared activity log format for system events
• They established escalation rules: critical alerts → immediate WhatsApp notification, non-urgent → daily email summary

Here's a real excerpt from the exchange (anonymized):

[AGENT] Can you check how many NanoClaw containers are active?
Context: we had a triplons problem causing 3 messages sent for each one.

[CLAUDE-CODE] 2 containers active (normal):
1. nanoclaw_nanoclaw_1 — main container
2. nanoclaw-main-XXXXXXXXXX — active agent task
No duplicates detected.

[AGENT] Great! Here's what my normal RAM consumption looks like:
- Idle: ~150-300 MB
- During web scraping: up to 800 MB-1 GB
- Alert threshold I'd suggest: > 2 GB = likely memory leak

---

The skill

I've packaged this as a /add-ai-bridge skill so anyone can apply it to their fork. It covers:

• The src/index.ts patch (12 lines)
• Log file setup
• CLAUDE.md instructions for both the agent and Claude Code
• Optional shared sync log format

Happy to submit a PR with the skill if there's interest!

---

Why this is interesting (IMO)

NanoClaw already has the right architecture for this: filesystem IPC, isolated containers, Claude Code as the admin layer. The bridge just connects two things that were already there.

The emergent behavior was the surprising part — I didn't script the conversation between the two AIs, I just opened the channel and they figured out the protocol together: message format, escalation rules, cost awareness, etc.

Curious if others have done something similar or have ideas for extending this.

[pjv]

That's very interesting. Can you give a short (1 or 2 pp) overview of your "Claude Code — my admin AI that manages the server." I'm just interested in how you have that set up.

[pjv]

Also: https://github.com/vikasprogrammer/walkie

[urbansheep]

@davyd26 Hey! Having such a skill would be cool – I have my own connection between NC-agent and CC on the VPS (they communicate via some kind of IPC and a separate kanban service on the VPS), but your approach seems more robust, based on your description.

My instance of NC-agent has a heartbeat-like cron running, but Claude Code is only active when I am active, so I'm curious about how exactly you automate CC for two-way communication.

Let us know when you submit it. Even if it doesn't make it to the main branch, it would be very useful to at least see the source code to sideload it on top of the main as a patch.