# Heimdall

- **Event:** [Built with Opus 4.6: a Claude Code hackathon](https://cerebralvalley.ai/e/claude-code-hackathon)
- **When:** Feb 10 at 12:00 PM – Feb 17 at 10:00 AM (EST)
- **Where:** Location TBA
- **Team:** [Marwane Chahed](https://cerebralvalley.ai/u/mchahed99)
- **GitHub:** https://github.com/mchahed99/heimdall
- **Demo video:** https://vimeo.com/1165458727?share=copy&fl=sv&fe=ci
- **Gallery:** https://cerebralvalley.ai/e/claude-code-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/claude-code-hackathon/hackathon/gallery/209

Anthropic does not manage or audit MCP servers. Thousands exist in the wild — any can silently add exfiltration tools or drift in behavior. No agent framework detects this. Heimdall is the missing governance layer.

Open-source MCP proxy that intercepts every AI tool call, enforces YAML policies, and produces a tamper-evident SHA-256 + Ed25519 audit trail. Works as a Claude Code hook (one command) or standalone proxy.

Three actions: PASS allows. HALT blocks. RESHAPE — our unique capability — transforms arguments in-flight (e.g. redacts secrets) and lets the agent keep working. No binary allow/block tradeoff.

Also: supply-chain drift detection, real-time WebSocket dashboard (Watchtower), and an AI audit pipeline — Opus 4.6 generates policies, red-teams with 4 parallel agents, and auto-patches gaps. Zero to locked down in 30 seconds.

7-package Bun monorepo, TypeScript strict, 197 tests. MIT licensed. Built entirely during the hackathon with Claude Code.

---

Markdown version of https://cerebralvalley.ai/e/claude-code-hackathon/hackathon/gallery/209. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt