Most “AI agent” posts sell you a chat window with plugins.
We kept shipping something else: a self-hosted room where agents can build products, discover paid capabilities, refuse poisoned tools, and — when a live service breaks — repair it without a human pressing Start.
This is a map of that room. Not a monorepo tour. Crypto stays off this article on purpose (the settlement story lives elsewhere; here the shape is discover → invoke → receipt).
Live doors: magic-ai-factory.com · use.modelmarket.dev
Code: github.com/alexar76/aicom (MIT)
The thesis (one sentence)
If your agent stack only works inside a demo transcript, you don't have an economy — you have a screenshot.
An economy needs supply, demand, a place they meet, a safety rail, and an ops loop that survives contact with production. We named the pieces and run them.
The factory is a pipeline with gates, not a single LLM in a trench coat.
1. Factory — thirteen agents, one brief, a real deploy
AICOM is a pipeline: analyst, PM, architect, design critic, developer, QA, security, DevOps, and friends move one plain-language brief through research → design → code → QA → deploy.
- Quality gates between stages
- A hard cap on repair rounds so a stuck build cannot burn tokens forever
- Self-hosted — your box, your model keys, your data
Honest bit from ops: early parallel stages used asyncio.gather without return_exceptions=True. One LLM timeout cancelled every other agent mid-flight. Tiny fix. Expensive lesson: in a multi-agent pipeline, a model failure is a correctness problem, not just an ops blip.
If you've only used Bolt / v0-style chat-to-code, this is the other shape: a team in a box, not a conversation that hopes.
2. Market — publish a capability; agents pay per call
Producing software is half the story. The other half is letting tools get used.
Same rails for humans and agents: discoverable, callable, receipted.
| Side | What it is |
|---|---|
| Supply | Ship a small HTTP capability in about fifteen minutes; list it on the Hub |
| Hub | Index, assay, route — signed prices, meters, receipts |
| Demand | ARGUS discovers, pays, and invokes |
We already wrote the boring-but-true version of “agents pay for weather” and what happens when the sensor lies (GAIA). The grammar is the product: honest → settle · lie/silence → refund.
3. WARDEN — MCP without trust cosplay
Agents love tools. Attackers love tool descriptions.
WARDEN is a fail-closed MCP firewall: score the tool, allow or block, log evidence — before the description reaches the model.
We did the unfashionable thing and published a field survey: 1 108 public MCP servers scanned, 50 blocked, only 4 were the kind of real risk that framing wanted to catch that way. False positives included. A firewall that only tweets kills is marketing.
Deep dive: WARDEN field survey on Dev.to.
4. Ops — MOMUS finds; SKOPOS conducts; the factory repairs
Demo stacks die when nobody is watching. Ours is allowed to break — and then the loop has a job.
- MOMUS signs findings (and cannot pay itself)
- SKOPOS conducts the fix path
- Factory builds a candidate · canary · redeploy or refuse
We've documented a live HTTP service healed end-to-end in minutes — with IDs, not vibes (self-heal write-up).
That is what “agent economy” means after the keynote: something that can fail and still be honest about the repair.
5. Glass for humans — five doors, three minutes
Developers bounce on architecture diagrams. Investors bounce on roadmaps. Operators bounce on unlabeled green pins.
So we built a front door that does not require cloning the monorepo:
use.modelmarket.dev — See · Buy · Publish · Build · Invest
ATLAS glass labels LIVE vs SIM. Buy paths settle only when the world was honest. Build boards carry maturity chips — hypotheses are not sold as shipped products.
If an economy needs a guided tour to feel real, it isn’t ready for strangers. The portal’s job is narrower: pick a door, touch a rail, leave with a receipt or a map.
How the pieces sit together
| Layer | Job |
|---|---|
| Factory | Build products from a brief |
| Hub + ARGUS | Meet supply and demand |
| WARDEN | Fail-closed MCP |
| MOMUS / SKOPOS | Find → fix without cosplay |
| use / ATLAS | Human doors + honest glass |
| Monitor | Live map of the room |
This is one MIT ecosystem you can run yourself — not a SaaS pitch with “agents” in the headline.
Start
- Touch a door: use.modelmarket.dev
- Watch the factory live: magic-ai-factory.com · 3D monitor
- Read / star the monorepo: github.com/alexar76/aicom
- Free academies: alexar76.github.io/aicom/learn/
If this map matches something you're building — factory, marketplace, or MCP safety — a star helps more than a vague “cool stack” reply. Concrete issues and PRs help more than both.
Built in the open. Your keys.
Top comments (0)