Turn Claude Code into a persistent AI agent — with a name, a brain that learns over time, and Discord access — running on your Mac.
One file. You feed it to Claude Code, and it walks you through building your own always-on AI agent, step by step, in about 15 minutes.
Claude Code gives you the wires. AgentWake gives you the brain.
AgentWake is a single guided Markdown file (agentwake.md). You open Claude Code, hand it the file, and it becomes an interactive setup wizard that builds you a persistent AI agent — one that:
- Has an identity — a name, values, a way of thinking, and a voice that stay consistent across every conversation.
- Has a brain that learns — a structured knowledge base (learnings, corrections, a journal, project notes, a profile of you) it reads on startup and updates as it goes, so it gets more useful over time instead of forgetting everything.
- Lives on Discord — talk to it from your phone, anywhere. It runs in a
tmuxsession on your Mac: always on, always reachable. - Improves itself — it records what it learns and where it was wrong, then applies that next time.
It's the open-source version of the setup behind Enzo, the persistent agent I run day to day.
Requirements: macOS 13+, a Claude Pro or Max subscription, and ~15 minutes.
- Download
agentwake.md(or clone this repo). - In a terminal, start Claude Code:
claude
- Point it at the guide:
Read agentwake.md and walk me through it. - Follow along. It installs the prerequisites (Homebrew, Bun, tmux, GitHub CLI, the Claude Code CLI), helps you create a Discord bot, builds your agent's brain, writes its
CLAUDE.md, and launches it in a persistent session.
You'll be chatting with your agent on Discord by the end.
The setup creates a structured brain your agent reads on every boot and updates in real time:
| File | What it holds |
|---|---|
Index |
Map of the brain + how the agent uses it |
Identity |
Who the agent is — values, voice, boundaries |
How I Think |
Reasoning approach + known blind spots |
Capabilities |
What it can/can't do, its tools, how it grows |
Learnings / Corrections |
What it's learned; where it was wrong |
Conversation Log / Journal |
Memory across sessions |
Active Projects / Project Ideas |
What it's working on |
<You> |
A profile of you — preferences, context |
…plus a permanent CLAUDE.md that ties it all together and a tmux launch setup so it stays running.
This gives your agent full access to the machine — files, terminal, browser, network — without approving each action. That autonomy is the point, but it means mistakes can have real consequences. Run it on a dedicated machine (not your primary workstation), keep backups, and only pair trusted people on Discord.
MIT — see LICENSE. Use it, fork it, build on it.
Built by Benji Banwart. I build AI agents — from personal setups like this one to production systems running real businesses. AgentWake is the open-source version of the setup I run myself.
→ benjibanwart.dev · @benjibanwart
If it's useful to you, a ⭐ helps others find it. Ideas or improvements? Open an issue or PR.