Forge

Agent supply chain security — code signing, SBOM, vulnerability scanning, and provenance attestation.

Forge ensures that every agent in the ecosystem is verified, signed, and scanned before it can be registered in Hive or authorized by Patroclus. It addresses OWASP Agentic AI Top 10 #5: Supply Chain Risks.

Governance:   Governance Hub · Patroclus · Argus · Forge
Runtime:      Hive · Relay · Miser
Enforcement:  Aegis · Sentiel · Cortex

Forge answers:

• "Was this agent code signed by a trusted publisher?"
• "Does this agent have known vulnerable dependencies?"
• "What's in this agent's software bill of materials?"
• "Who built this agent, from what source, when?"
• "Can we trust this agent with production access?"

• Ed25519 signatures on agent packages (tarballs, containers, binaries)
• Signing key management with key rotation
• Verification against trusted publisher registry
• Signature verification integrated into Hive agent registration

• Software Bill of Materials for every agent package
• Dependency tree with versions and licenses
• Formats: SPDX, CycloneDX
• Stored alongside signature for audit trail

• Scan agent dependencies against vulnerability databases
• Rust (cargo audit), Python (pip-audit / safety), Node (npm audit)
• Severity classification (critical/high/medium/low)
• Block registration if critical vulnerabilities found (configurable)

• SLSA-style provenance: who built it, from what source, when, how
• Build environment attestation (CI/CD pipeline, commit hash, builder identity)
• Provenance bound to signature — can't forge provenance without the key

• Verified publisher → higher trust → broader permissions
• Trust factors: code signed, SBOM clean, no critical CVEs, provenance verified
• Trust score feeds into Patroclus policy (min_trust_level constraint)
• Trust decay: score decreases if not re-scanned within threshold

cargo build --release
./target/release/forge serve --config config.toml
# API: http://localhost:8788

Hive calls Forge before accepting an agent registration:

# Verify agent package before registration
curl -X POST http://localhost:8788/api/verify \
  -d '{"package_path":"/path/to/agent.tar.gz","publisher":"trusted-org"}'
# Returns: {"verified":true,"trust_score":0.92,"sbom":{...},"vulnerabilities":[]}

Patroclus policies can reference Forge trust scores:

- name: high-trust-only
  actions: ["deploy"]
  resources: ["prod-*"]
  decision: allow
  min_trust_score: 0.8
  reason: Only high-trust agents can deploy to production

Forge is part of the AI governance ecosystem governed through Governance Hub:

Forge gates agent registration in Hive: before an agent can be deployed, Forge verifies its signature, generates an SBOM, scans for vulnerabilities, and calculates a trust score. Agents with critical vulnerabilities are blocked. Trust scores feed into Patroclus policies (e.g., only trust score > 0.8 can deploy to production).

Run the full ecosystem:

~/patroclus/scripts/start-ecosystem.sh start  # Starts all services

See the ecosystem documentation for the complete integration guide.

Early development.

MIT