Maia is an open-source framework for creating a personal AI agent that runs 24/7, communicates through Slack, remembers everything, and handles tasks autonomously. Simpler than OpenClaw, more transparent than anything else. Powered by Claude Code or Codex CLI.
Simple setup
Your agent runs on a VM (or local machine), listens for messages via Slack, and uses Claude Code or Codex CLI as its brain. Setup takes under 15 minutes.
Create a Hetzner Cloud server for ~$9/month. Run our hardening script to lock it down. Or use your own machine.
Install Claude Code or Codex CLI. Clone the repo. Paste the setup prompt into your agent — it configures itself.
Create a Slack app, add your tokens, start the bot. Your agent is now live — message it anytime.
Full-featured
Not just a chatbot — a full operating system for an autonomous AI assistant.
4-layer memory system — your agent remembers everything across sessions. Durable facts, daily notes, knowledge graph, and session notes with automatic recall. Self-improving feedback detection learns from your corrections.
Chat with your agent through Slack using Socket Mode. Full thread support, file sharing, live tool-use status updates, and multi-channel routing.
Spawn long-running tasks that run autonomously and report back when done. Research, code generation, data processing — all in the background.
Cron system powered by systemd timers. Morning summaries, inbox triage, memory consolidation — your agent works while you sleep.
Receive and act on external events — email, GitHub, payments. Your agent triages and responds automatically.
Trusted channel enforcement, prompt injection defense, VM hardening. Your agent knows what to trust and what to ignore.
Transparent
File-based everything. No databases. Fully inspectable and debuggable.
Your VM / Machine ├── CLAUDE.md or AGENTS.md ← Identity, rules, skills registry ├── .claude/ or .codex/ ← Hooks, skills, settings ├── .agent/ ← Scripts, cron, workers │ ├── scripts/ │ │ ├── hooks/ ← Lifecycle hooks │ │ ├── auto-recall.py ← Memory search on every message │ │ ├── feedback-detector.py ← Correction detection + queue │ │ ├── spawn-worker.sh ← Background task spawner │ │ └── heartbeat.sh ← Worker health monitor │ ├── cron/ ← Scheduled job system │ └── workers/ ← Worker state tracking ├── memory/ ← Daily notes (auto-generated) ├── life/ ← Knowledge graph (PARA method) ├── sessions/ ← Conversation session notes └── slack-bot/ ← Slack Socket Mode bot
Honest comparison
OpenClaw got 160k GitHub stars and a lot of hype. Here's why we think Maia is the better foundation for a personal AI assistant.
File-based everything. Markdown files, shell scripts, plain JSON. Fully inspectable — open a file, read it, done.
5-component system (Gateway, Brain, Memory, Skills, Heartbeat) with its own abstractions. More to learn, more to break.
Uses Claude Code or Codex CLI directly — the same tools developers already know. No wrapper, no custom agent loop. You get the full power of the CLI.
Custom ReAct reasoning loop wrapping LLM API calls. Another layer between you and the model.
Copy-paste a setup prompt. The agent configures itself in ~15 minutes. No framework to learn.
Install the framework, learn the skill format, configure channels, set up the Gateway. Steeper learning curve.
Something wrong? Read the file. Every piece of state is a file on disk — memory, config, logs, worker output. grep and you're done.
Debugging means tracing through the ReAct loop, Gateway routing, and skill execution. More moving parts = more places things go wrong.
Actively maintained by the creator. No VC funding, no acquisition risk. MIT license — it's yours forever.
Creator joined OpenAI in Feb 2026. Project handed to an open-source foundation. Future direction uncertain.
Bring your existing Claude Code or Codex CLI subscription. No additional framework costs. Starts at ~$30/mo total.
Bring your own API key, but you're paying per-token through their agent loop — less efficient than native CLI tools.
OpenClaw is a good project. But if you want something simple, transparent, and built to last — Maia is the better choice.
Proof it works
This web app — the design, the code, the deployment — was built and shipped autonomously by a Maia-powered agent, directed entirely through Slack messages. No IDE, no manual coding. Just conversation.
It's also updated weekly, automatically. As the creator improves his own agent's infrastructure, a scheduled job reviews the changes and syncs them to this open-source project — no human intervention required.
Follow the step-by-step guide. You'll have a working AI assistant in under 15 minutes.
Start the Guide