OpenAgentAudit

Open evidence format and Cloudflare-native audit toolkit for enterprise AI agents.

Trustavo — Trust, voiced with authority. The production deployment at trustavo.com is the reference implementation of OpenAgentAudit. In audit, evidence only counts when it is trusted; Trustavo exists to make that trust legible.

Agent logs are not audit evidence. OpenAgentAudit turns tool calls, policy decisions, human approvals, benchmark results, training manifests, and runtime traces into defensible technical evidence reports.

Live deployment

trustavo.com — production deployment on Cloudflare Workers.

Install

npm install @openagentaudit/schema @openagentaudit/core @openagentaudit/adapters

import { validate, computeRiskScore, renderReport } from '@openagentaudit/core';
import { validateEvents } from '@openagentaudit/schema';
import { aepV0_2, otel, langfuse, langsmith } from '@openagentaudit/adapters';

// Parse and validate canonical events
const { valid: events } = validateEvents(raw);
const { errors, crypto_summary } = await validate(events);

// Run the full audit pipeline
const score = await computeRiskScore(events, runId);
const bundle = await renderReport(events, findings, score);
// bundle.html, bundle.markdown, bundle.json, bundle.csv

What it does

• Validate agent evidence records against a canonical schema.
• Reconstruct tool and permission inventory from runtime traces.
• Detect policy boundary violations and excessive agency.
• Audit benchmark claims with paired statistics (McNemar, Wilson CI).
• Check contamination risk with CPU-first chunked algorithms.
• Monitor behavioral drift over time.
• Render training-run audit reports from trace-pipeline evidence bundles.
• Map evidence to OWASP Agentic Top 10, NIST AI RMF, ISO/IEC 42001, and EU AI Act Annex IV technical documentation needs.
• Run as a TypeScript/Bun CLI or as a Cloudflare-native service.

What it does not do

• It does not provide legal advice.
• It does not certify regulatory compliance.
• It does not require GPU.
• It does not require Python, scipy, or a traditional backend server.
• It does not train models or run benchmarks directly.
• It does not replace observability tools — it consumes their traces and produces audit evidence.

Where it sits in the WasmAgent ecosystem

wasmagent-js (runtime, SDK, AEP emitter)
        │
        ├─── bscode (real coding-agent workload)  ──┐
        └─── erp-agent (ERP workload, planned)    ──┤
                                                    │ AEP JSONL
                                                    │ (signed runtime evidence)
                                                    ▼
                        ┌────────────────────────────────────────────────────┐
                        ▼                                                    ▼
              trace-pipeline                                       open-agent-audit
              Measurement Trust                                    (this repo)
              Evidence Admission Gate                              ─────────────────
              Training Audit Backend                               audit reports
                │                                                  regulatory maps
                │  AgentTrustScore                                 benchmark claims
                │  + training evidence ──────────────────────────► evidence bundles
                │
                ├── SFT / DPO datasets (gated — regression gate required)
                └── ADAPTER_CARD.md   (promote / hold / reject)

External Observability (OTel, Langfuse, LangSmith) ─── via adapters ──► open-agent-audit

OpenAgentAudit is the reporting and evidence layer of the WasmAgent ecosystem. trace-pipeline decides whether a benchmark claim is statistically credible and whether a training run is auditable; OpenAgentAudit turns that evidence into enterprise-readable reports. Both consume AEP records; they serve different downstream customers.

Cloudflare-native by design

OpenAgentAudit's reference deployment is Cloudflare-only:

• Cloudflare Workers — API and audit engine
• Cloudflare Pages — Dashboard
• Cloudflare R2 — traces, artifacts, reports
• Cloudflare D1 — run / finding / evidence metadata
• Cloudflare Queues — chunked async jobs
• Cloudflare Durable Objects — per-run coordination
• Cloudflare Workflows — durable multi-step orchestration
• Cloudflare Browser Run — HTML → PDF
• Cloudflare Containers — optional heavy enterprise jobs

No external VPS, Cloud Run, or Kubernetes is required.

Status

Phase 2 — Active implementation. The specification, schemas, and regulatory profiles are complete. TypeScript implementation packages are being built against the current open-agent-audit/v0.1 schema.

Component	Status
spec/versions/v0.1/SPEC.md	draft
schemas/v0.1/*.schema.json	draft
profiles/*.yaml	draft
packages/schema	implemented — Zod runtime validation
packages/core	implemented — all engines operational
packages/adapters	implemented — AEP v0.2, bscode, OTel GenAI, Langfuse, LangSmith; 444 tests
packages/cli	implemented — 7 commands including from-aep, from-bscode
packages/worker	implemented — REST API, deployed at trustavo.com
packages/dashboard	implemented — React SPA deployed at trustavo.com
npm packages	published — @openagentaudit/schema, @openagentaudit/core, @openagentaudit/adapters

Documents you should read first

1. CONSTRAINTS.md — project rules; every contributor must read this.
2. spec/SPEC.md — canonical evidence specification.
3. docs/architecture.md — system architecture.
4. docs/cloudflare-native.md — deployment model.
5. docs/relationship-to-wasmagent.md — how this fits with wasmagent-js, bscode, trace-pipeline.
6. docs/schema-versioning.md — versioning policy.
7. docs/regulatory-disclaimer.md — what we do and do not claim.
8. docs/compliance-coverage-report.md — per-framework coverage depth and breadth (OWASP / EU AI Act / NIST AI RMF / ISO 42001), upgrade paths, and ceiling analysis.
9. docs/competitive-landscape.md — market analysis: how OpenAgentAudit compares to ATR, VerifyWise, Credo AI, and other tools in the AI agent audit space.

Disclaimer

OpenAgentAudit produces technical evidence that may support selected regulatory documentation requirements. It does not constitute legal advice, regulatory certification, or a determination of compliance. Regulatory interpretations evolve; users are responsible for their own compliance posture. See docs/regulatory-disclaimer.md.

Acknowledgements

OpenAgentAudit is part of the WasmAgent ecosystem.

Runtime dependencies: Zod (MIT), React (MIT), Tailwind CSS (MIT).

See NOTICE for full third-party attributions.

License

Apache License 2.0 — see LICENSE.