Agent Fleet
A generated swarm of test agents that red-teams MCP servers
- Status
- Live demo
- Year
- 2026
- Rooms
- AI & Agents
Free host: the first load takes about a minute to wake.
Problem
Every MCP tool exposed to an agent is an API untrusted input can reach, and nobody hand-writes hundreds of adversarial tests for ten tools.
Approach
Agent Fleet reads a server's live tool manifest, generates 305 single-purpose worker specs from 12 squad templates, runs them with bounded concurrency, has a judge grade every failure, and files tickets for the real bugs.
Engineering decisions
- Worker prompts are generated from templates, never hand-written — the design is 322 roles (305 workers plus orchestration), dispatched through a semaphore pool, not 322 models running at once.
- The judge re-reviews failures and a sample of passes to catch false negatives.
- Ships with a deliberately vulnerable demo server whose three planted bugs it finds without hints.
What it does
- Manifest-driven agent factory
- Streamed dashboard over SSE
- Severity judging with rationale
- GitHub issues & Slack summary
Stack
My part
All 24 commits in the repository are mine.
Limits, stated plainly
- Demonstrated against the bundled ShopFlow demo server.
- Free hosting has a cold start.
Evidence
- Agent-Fleet README
Screenshots
Agent Fleet's live landing page