SHEPHERD
Self-extending AI agents with WASM sandbox. Extensions are generated, scanned, compiled, and executed at runtime — all within a security perimeter enforced by gRPC contracts, capability tokens, and an HMAC-chained audit log.
How it works
- 01User sends natural-language message via Telegram
- 02Claude classifies intent, generates Go source code
- 03SAST scans code (Go AST) — blocks dangerous patterns
- 04Compile to WebAssembly (wasip1 target)
- 05Execute in wazero sandbox with capability token
- 06Return structured result, audit every event
- 07Cache extension for instant reuse (~3s vs ~15s)
Three-layer security
Layer 1: SAST
Go AST blocks forbidden imports, calls, env access, hardcoded secrets.
Layer 2: WASM
Zero syscalls, zero sockets, isolated linear memory. Escape-proof by design.
Layer 3: Policy
Domain allowlist (compiled), HMAC tokens, rate limiting, HMAC-chained audit.
Built and verified (Phase 1)
Roadmap
Phase 1 — Live
- WASM Sandbox
- SAST + Tokens
- Audit + gRPC
Phase 2 — Q2 2026
- Versioning
- Scratchpad
- mTLS
Phase 3 — Q3 2026
- Multi-User
- OPA Policies
- NIS2 Mapping
Phase 4 — Q4 2026
- Web UI
- 1000+ Scale
- Multi-Tenant
Why it matters
Risk Reduction
Every agent action is sandboxed, scanned, and audit-logged before it touches production.
Compliance Path
HMAC-chained audit trail and capability tokens create a verifiable trust chain.
Proven Tech
71 tests, Phase 1 live, real Telegram traffic — not a prototype.
Low Overhead
Single Go binary, Docker Compose up, no Kubernetes required.