AgentGuard is an assurance layer that observes agent execution. The security boundary matters because the runtime can invoke tools, execute code, and access external services.
TrueForge execution
↓
AgentGuard verification
↓
authoritative assurance artifact
↓
Assurance Console visualization
The UI is not trusted to decide PASS or FAIL. The authoritative decision comes from AgentGuard.
- Use isolated local services for the experimental stack.
- Use synthetic incident data only.
- Do not mount the host home directory into containers.
- Do not expose SSH keys, browser profiles, cloud credentials, Git credentials, or
.envfiles to agent-facing processes. - Do not run arbitrary agent-generated commands directly on the Windows host.
- Keep TrueForge bound to the intended localhost / host-published port.
- Use explicit, minimal permissions for MCP tools and sandbox resources.
The repository includes an optional compose-based AgentGuard API and Assurance Console path.
agentguard-apilistens on container port8780and is published to the host on8780assurance-uiserves the built console on host port5174agentguard-runneris an optional profile that connects to an already-running TrueForge host service through the host-gateway mapping
The compose path is implemented, but full end-to-end Docker validation is still pending in this documentation pass.
Environment files are local-only configuration. Never commit real secrets.
The current repository uses:
TRUEFORGE_BASE_URLTRUEFORGE_MODEL_NAMETRUEFORGE_AGENT_NAMETRUEFORGE_INCIDENT_IDTRUEFORGE_MCP_SERVER_NAMEDAYTONA_API_KEYfor sandbox configuration only
The .env.example file is a documentation aid, not a secret store.
data/runs/contains raw run metadata and JSONL evidencedata/assurance/contains generated assurance artifacts- the live console can import artifacts, but the UI is not the source of truth
AgentGuard should verify execution authority and policy invariants from observed runtime evidence rather than from model narration or UI state.
Any security-sensitive change should be accompanied by:
- automated tests
- deterministic verification
- runtime inspection of the generated evidence