nanoclaw-discord — NanoClaw preconfigured for Discord and Docker. One process, one channel stack, no config matrix.
Message Claude from Discord. Agents run in Docker. Small codebase, secure by isolation.
NanoClaw is a minimal, understandable Claude assistant: one Node process, agents in containers, no framework bloat. nanoclaw-discord is that same stack, shipped for a single main use case:
- Discord as the only messaging channel.
- Docker as the only container runtime (macOS and Linux).
No “choose your runtime” or “add Discord via skill” — it’s Discord + Docker by default. You get a small, auditable codebase that does one thing clearly.
git clone https://github.com/gavrielc/nanoclaw.git
cd nanoclaw
claudeThen run /setup. Claude Code will walk you through: dependencies, Discord bot token, Claude auth, Docker image, and registering your first channel.
One stack. Discord for I/O, Docker for agents. No optional runtimes or channel providers to wire up.
Small enough to understand. One process, a handful of files. No microservices, no message queues. Have Claude Code walk you through it.
Secure by isolation. Agents run in Docker containers with explicit mounts. They only see what you mount. Bash inside the container is safe.
Built for one user. Fork it and have Claude Code adapt it to your workflow. Customization = code changes, not config files.
AI-native. No install wizard; Claude Code guides setup. No dashboard; ask Claude what’s happening. No debug UI; describe the problem, Claude fixes it.
Best harness, best model. Runs on Claude Agent SDK (Claude Code). The harness matters; this one is minimal and transparent.
- Discord I/O — Register channels in
data/registered_groups.json; only those channels trigger the agent. Trigger word (default!nano) in each channel. - Isolated group context — Each channel has its own
CLAUDE.mdmemory and filesystem; each run is in its own container with only that group’s data mounted. - Main channel — One Discord channel is “main” (admin): register other channels, manage scheduled tasks, optional extra mounts.
- Scheduled tasks — Recurring or one-time jobs that run Claude and can message back via Discord.
- Web access — Search and fetch; agent can use browser automation (agent-browser) when you add the skill.
- Optional skills — Add Gmail (
/add-gmail), more channels, or custom skills via.mdincontainer/skills/and CLAUDE.md.
In any registered Discord channel, use the trigger (e.g. !nano):
!nano send a summary of the sales pipeline every weekday at 9am
!nano every Monday at 8am, compile AI news from Hacker News and message me a briefing
!nano what skills do you have?
From the main channel you can also:
!nano list all scheduled tasks
!nano pause the Monday briefing task
!nano register the #general channel
See docs/SETUP-GROUP.md for registering channels and docs/SETUP-GROUP.md#optional-choosing-a-model for setting CLAUDE_MODEL in .env.
Tell Claude Code what you want; the codebase is small enough to change safely:
- “Change the trigger word to !bob”
- “Remember to keep responses shorter”
- “Add a skill for Notion” (and add
container/skills/notion-api.md+ a bullet ingroups/main/CLAUDE.md)
Or run /customize for guided changes.
- macOS or Linux
- Node.js 20+
- Claude Code (for setup and editing)
- Docker
Discord (discord.js) → MessageCreate events → Host (Node) → Docker container (Claude Agent SDK) → Response
Single Node process on the host: Discord connection, routing, scheduler, IPC. Each message (or scheduled task) spawns a short-lived Docker container with the group’s mounts; the container runs Claude Agent SDK and exits. No daemons, no queues.
Key files:
| File | Purpose |
|---|---|
src/index.ts |
Discord connection, message routing, IPC |
src/container-runner.ts |
Builds mounts, spawns docker run |
src/task-scheduler.ts |
Scheduled tasks |
src/db.ts |
SQLite (tasks, run history) |
groups/*/CLAUDE.md |
Per-channel memory and instructions |
data/registered_groups.json |
Registered Discord channel IDs |
Why Discord only?
This variant is Discord + Docker only. To add another channel, use a skill or fork and add it.
Why Docker only?
So the same setup runs on macOS and Linux with one runtime. No Apple Container or other backends.
Can I run this on Linux?
Yes. Install Docker and Node, clone, run /setup with Claude Code.
Is this secure?
Agents run in containers with explicit mounts; they don’t see your full filesystem or .env (only a small allowlist of vars is passed in). See docs/SECURITY.md for the security model.
How do I debug issues?
Ask Claude Code: “Why didn’t the last message get a response?” “What’s in the latest container log?” Or run /debug for guided checks.
What’s the difference from upstream NanoClaw?
Upstream may support other runtimes/channels via skills. This setup locks to Discord + Docker for a simpler, single-use-case.
Prefer skills over in-repo features.
If you want Telegram or Slack, contribute a skill (e.g. .claude/skills/add-telegram/SKILL.md) that transforms a clone to add that channel. Keep the base as Discord + Docker.
MIT