This is what my "team" looks like. 7 AI agents, one dashboard, no employees.

I want to show you something instead of just describing it.
This is my operations dashboard on a normal week:
![APES OS Operations Report — 32 agent runs over 7 days, 66% success rate, top agents: Koko, Turing, Atlas, Caesar]
32 runs in 7 days. 66% success rate. Koko the content writer leads with 9 runs. Turing the developer at 7. Atlas the researcher at 6.
I didn't do any of that work.

What you're looking at
This is APES OS — a local multi-agent AI system I built on top of OpenClaw. It runs entirely on my machine. Seven agents, each with a specific job, each with their own Discord channel and workspace folder.
Here's the agent status page — this shows every agent in real time:
![APES OS Agent Status — all 7 agents showing green CONNECTED, with last task summaries and model info]

All seven showing green. Connected, idle, ready.
The ones with "No data yet" are Main and APES — they only activate when I trigger an orchestration run or a full pipeline. Atlas, Koko, Caesar, Turing, and Zuri handle the day-to-day requests that come through Discord.
The models you can see — openrouter/stepfun/step-3.5-flash and gemini-3-flash — are both free tier. That's the whole stack.

The part I use most
The kanban board. This is where actual work gets tracked:
![APES OS Task Board — 12 tasks in Done column including content creation, research, and Instagram posts]

Twelve tasks in Done. Zero in To Do or Doing. The dashboard shows "Outstanding — you crushed it" which is a little embarrassing but accurate for that day.
The tasks you can see in the Done column are real ones from that week:

create beautiful image post for vizora launch (assigned to Zuri)
write instagram post of vizora launch at 25/3/26 (Zuri again)
research best ai models for coding (Atlas)
research for a database related saas present in market (Atlas)

All of them completed, logged to Supabase, visible in the activity feed and charts.

How a task actually moves through the system
I type something in Discord — whichever channel matches the agent I want:
Research the best database solutions for
a bootstrapped SaaS in 2025. Focus on cost,
scalability, and what solo founders actually use.
That goes into #atlas-research. The Discord router picks it up, maps the channel ID to Atlas, spawns an OpenClaw CLI process with the message attached, and Atlas runs it.
When Atlas is done, the response comes back in Discord. Simultaneously, the completion gets logged to the agent_logs table in Supabase with the model used, duration, and a summary of the output.
The dashboard pulls from that table. The chart updates. The activity feed shows the run.
The whole thing happens without me touching anything after sending the message.

The pipeline that does the heavy lifting
For content work, I run a four-step chain:
Run full pipeline on [topic]:
Atlas researches → Koko writes the blog →
Zuri makes social posts → Caesar builds
a 30-day promotion plan
One message. About twenty minutes. I come back to four separate outputs.
I still review everything — especially Zuri's social content which needs editing — but the starting point is already 70-80% there. For research and blog drafts it's closer to 90%.

What the 34% failure rate means
I showed the success rate screenshot on purpose. 66% means roughly one in three runs fails.
Usually it's a free model hitting a rate limit mid-run. Sometimes the fallback chain catches it and retries with a different model. Sometimes it just fails and I re-send the message.
It's not a polished product. It's infrastructure I built for myself on weekends. The failure rate is annoying but the overall output volume more than makes up for it — even at 66%, I'm getting more done than I was doing everything manually.

The technical stack if you want to replicate this

OpenClaw — open source, handles the agent gateway and CLI
Discord.js — custom router that bridges messages to agents
Supabase — free tier, stores tasks and agent logs
OpenRouter — free model tier, stepfun flash as primary with fallbacks
Node.js — serves the dashboard locally on port 3000
PowerShell — one startup script launches all three services

Everything runs locally. No cloud hosting cost. No subscription.

If you want to build something like this
The full setup took me a weekend to get working — OpenClaw config, Discord bot permissions, Supabase schema, the startup script, all seven agent SOUL files. It's not complicated but there are a lot of moving pieces and a lot of ways to get stuck.
I documented the whole thing — every step, every config, every error I hit with the exact fix — in a Notion template with a PDF version and all the agent SOUL files pre-written.
It's at nexflowai.gumroad.com/l/npzufj if you want to skip the trial and error.
If you'd rather poke around first, OpenClaw's docs are at docs.openclaw.ai. The source is on GitHub.
Happy to answer questions about any part of how this runs — the Discord routing, the Supabase schema, the OpenClaw config, whatever you're stuck on.