Trust infrastructure for the agent economy.
Software is increasingly consumed by agents, not humans. Swarm Proof builds the open-source tooling that makes that safe — before real agents (and real money) arrive.
Seven projects, one thesis: what happens when autonomous agents meet real systems and real money — and how do we make it safe? Each earns its place alone; together they're one agent-reliability stack.
| Project | What it does |
|---|---|
| 🐎 stampede | The wind tunnel for the agent economy — point a herd of realistic (and adversarial) agents at your system before real ones arrive |
| 🌍 mockworld | A synthetic internet for agents — fake Stripe, Gmail, exchange, and more as instant MCP servers |
| 🔬 mcp-probe | The CI quality suite for MCP servers — lint, contract-test, benchmark, load-test before you ship |
| 💸 costbomb | Denial-of-wallet fuzzing — find the inputs that make your agent spend $500 to answer a $0.05 question |
| 🎯 exactly-once | Idempotency middleware so agent side-effects — payments, emails, onchain txs — fire exactly once |
| 📓 agent-postmortems | A structured incident database + post-mortem standard for real agent failures in the wild |
| 🗺️ awesome-agent-reliability | The curated map of the field |
Applied case study: cairn-protocol — state-recovery for agents, the first system these tools are proven against.
- Building an MCP server? → mcp-probe scores it before you publish.
- Agents that move money? → drop in exactly-once and fuzz with costbomb.
- Shipping anything agent-facing? → run stampede against it and read the Agent Readiness Report.
- Just mapping the space? → start at awesome-agent-reliability.
Provider-agnostic (any OpenAI-compatible endpoint + Anthropic SDK, local-model friendly) · honest over impressive (we document the boundaries of our guarantees) · watchable & reproducible (seedable runs, screenshot-worthy reports) · Apache-2.0 everywhere.
Contributions welcome — every repo ships with seeded good-first-issues.
