Okay. Let's dive into another trending repo in the AI agent storm / agents-working-together / multi-agent-orchestration space. This time they're claiming a little bit of AGI in the name which but yeah it's punchy and rhetorical so I won't punish them for it.
As always no shade or hate to the creator, thank you for building things. Not really interested in credentials just gonna poke around and do my thing.
TL:DR
The 8-bit pixel art version of "we have Claude Code at home." Honestly? Kinda fun and cute. Good for people who don't like getting stuff done quickly.
Verdict: Butter smooth installation. Eye candy. Cool to poke at, not keeping it installed.
| What it is | A message queue + routing layer that wraps Claude Code CLI into "agents" with a web dashboard and Discord/Telegram/WhatsApp channels |
| Stars | 3.4k |
| Make No Mistakes Take | Fast setup, small security prob, solves a problem almost nobody has |
| Who it's for | Devs who want multiple Claude instances coordinated across chat channels |
TinyAGI
Multi-agent, Multi-team, Multi-channel, 24/7 AI assistant
The Stats
3.4k stars — v0.0.20 — ~9,500 lines of TypeScript backend + ~15,000 lines of Next.js frontend — dependencies are lean
Let's Get This Out Of The Way
What this actually is? Another trenchcoat wrapper around Claude Code. Each "agent" is literally just a claude CLI process running in its own folder. "Multi-agent collaboration" is string parsing for bracket tags that re-queue messages to the next agent. There is shared context + a directed message to the next agent.
There's no shared reasoning, no consensus, no agents learning from each other. It's separate claude sessions with a router in front.
But the thing its nice because it just works out of the box. No annoying setup. It's like a lite version of what OpenClaw wishes it was.
How It Actually Works
curl -fsSL https://raw.githubusercontent.com/TinyAGI/tinyagi/main/scripts/install.sh | bash tinyagi
That's it. You don't need to configure agents or skills or whatever they come out of the box. And just like you would use Claude agent teams (because yeah, it's a wrapper for Claude agent teams), you ask it to do a task and it spins up the team and does it.
The cool thing and kind of their claim to fame is this SQLite message queuing system. WAL mode, atomic transactions, retry logic with a dead-letter queue for failed messages. You can talk to each one of the agents while they're doing stuff, which
they handle better than Claude agent teams does raw.
Messages come in from Discord or Telegram or WhatsApp or the web UI, hit the queue, get routed to the right agent, and responses come back. It actually works.
And then there's the pixel office.
TinyOffice is their web dashboard and it has this little pixel art office simulation where your agents walk around as sprites. It's some real fun eye candy... A busy dev is not spending any time here. but yeah..
The Security Problem Nobody's Talking About
This is where it gets spicy. I had 5 Claude critics tear through the source code and the security findings are... not great.
The local API running on localhost:3777 has zero authentication. Wide open CORS. Any JavaScript running in any tab in your browser can call fetch('http://localhost:3777/api/settings') and get back all your API keys in plaintext. Anthropic key, OpenAI key, Discord bot token — all of it. Just sitting there.
It gets worse. The hosted dashboard a office.tinyagicompany.com connects to that same unauthenticated localhost API. So you're running an API with no locks on the door while a third-party web app reaches into it.
Not meant to be malicious. Just built for local dev convenience without thinking about the attack vectors.
After Using It
Setup — genuinely fast. Two minutes from zero to working. No complaints.
The default agent responded, ran tools, did its thing. It works. You can spin up multiple agents, route messages with @agent prefixes, and the queue keeps everything in order.
But there's really not much else to say. It works smoothly for what it does but what it does is less than what the README sells you. The "multi-team collaboration" is bracket-tag parsing. The "parallel processing" is just different agents running
concurrently (which... yeah, that's what separate processes do). The "persistent sessions" are just Claude Code's own built-in session persistence. TinyAGI doesn't implement that it just grabs it from Claude.
The Verdict
Will I use it again? No. It's fun but not for people who ship fast.
Report Card
| Category | Rating |
|---|---|
| Overhype Meter | 🔴 High — "AGI" in the name, "multi-agent collaboration" is string parsing, README oversells |
| Setup Pain | 🟢 Low — genuinely fast, two minutes to working, best part of the whole thing |
| Slop Level | 🟡 Medium — real code and clean architecture under the hood, but security gaps |
| Fun Factor | 🟢 High — pixel office is adorable, watching agents spin up is genuinely entertaining |
| Usefulness | 🔴 Low — the queue and channel routing work, but most people don't need this |
| Would I Keep It Installed | 🔴 No — not with that security posture, and the pixel office isn't enough to save it |