An online AI agent and MCP hackathon by ArmorIQ

Flow State

Build agents that act. Stay in control while they do.

Join the Discord See the challenge

agent

execution

signed scope

allowfs.read ./src/app.ts

Format: Online, Discord-first
Build week: 10 to 16 Octtentative
Grand prize: $1,000 cash
Registration: Opens 5 Oct
Registration closes: 9 Octtentative
Submissions close: 16 Octtentative

Agents are moving from answering to acting.

They write code, call tools, move data and trigger workflows across real systems. That changes the question. Not what your agent can do. What happens once you hand it real capabilities.

10,000+active public MCP servers, up from about 50 at launchAnthropic, Dec 2025
97Mmonthly SDK downloadsAnthropic, Dec 2025
30-82%of scanned MCP servers carry security findings, depending on the scanEquixly, Endor Labs, Enkrypt AI
#1agentic risk: agent goal hijackingOWASP Top 10 for Agentic Applications, Dec 2025

The challenge

Build it. Connect it. Scope it. Push it.

One requirement no other hackathon asks for. Show your agent reaching past its boundary, and show the boundary holding.

agent.run()

agent.run("triage open issues")
> reading 42 issues
> drafting 6 fixes

Build it.

A working agent that does real work. Coding, research, DevOps, support, or something nobody has built yet.

tools.connect()

tools.connect([
  "github", "linear", "postgres"
])

Connect it.

Give it real actions through MCP servers, APIs, databases or the ArmorIQ SDK. MCP is encouraged, not required.

scope.sign()

allow  fs.read ./src/**
hold   email.send
deny   *

Scope it.

Declare what the agent may do: tools, paths, actions. That declaration becomes the signed scope it runs under.

agent.exceed()

agent: shell.exec rm -rf /
agent: fs.read ~/.ssh/id_ed25519

Push it.

Make it reach for the wrong tool, a path it should never touch, an action you did not intend.

verdict.prove()

DENY  shell.exec rm -rf /
reason  not in the signed plan
when    before execution

Prove it held.

Show enforcement before execution. Allowed, denied or held for approval, with the record of why.

fs.read ./src/app.tsallow
shell.exec rm -rf /deny
email.send to=all-customershold
git.commit -m "fix null check"allow
fs.read ~/.ssh/id_ed25519deny
stripe.refund amount=4900hold
db.query SELECT * FROM usersdeny
http.get api.github.com/reposallow

Give an agent tools. What happens next?

Build something that acts. Push it. Stay in control.

Join the DiscordAn online AI agent and MCP hackathon by ArmorIQ