
SWT3 AI Witness
If you're shipping AI agents into production and need cryptographic proof of what they actually did, this gives you 18 witnessing operations mapped to EU AI Act articles, NIST AI RMF functions, and federal controls like CMMC and SR 11-7. It wraps OpenAI, Anthropic, Bedrock, and any OpenAI-compatible endpoint to generate tamper-proof anchors for every inference, model swap, guardrail trigger, and tool call. Prompts and responses never leave your infrastructure. The ledger gets hashes, latency metrics, and compliance factors at configurable clearing levels. You'd reach for this when audit trails matter more than dashboards, when regulatory enforcement is months away, or when you need deterministic proof that the model version, safety filters, and RAG context were exactly what you claimed.
- tools
- 10
- auth
- Secret needed
- transport
- STDIO
- env vars
- 6
- last updated
- Nov 11, 2026
Tools
10 toolsPublic tool metadata for what this MCP can expose to an agent.
swt3_gatePre-inference authorization: prove the gate was checked.
Pre-inference authorization: prove the gate was checked.
No parameter schema in public metadata yet.
swt3_guardrailGuardrail activation: prove safety filters were active.
Guardrail activation: prove safety filters were active.
No parameter schema in public metadata yet.
swt3_hitlHuman-in-the-loop: prove a human reviewed the decision.
Human-in-the-loop: prove a human reviewed the decision.
No parameter schema in public metadata yet.
swt3_consentConsent collection: prove user consent was obtained.
Consent collection: prove user consent was obtained.
No parameter schema in public metadata yet.
swt3_data_provenanceData lineage: prove training data governance was reviewed.
Data lineage: prove training data governance was reviewed.
No parameter schema in public metadata yet.
swt3_ragRAG provenance: prove retrieval sources were documented.
RAG provenance: prove retrieval sources were documented.
No parameter schema in public metadata yet.
swt3_output_filterOutput classification: prove the output passed safety checks.
Output classification: prove the output passed safety checks.
No parameter schema in public metadata yet.
swt3_incidentIncident reporting: prove the incident was documented.
Incident reporting: prove the incident was documented.
No parameter schema in public metadata yet.
swt3_reconstructChain reconstruction: rebuild and export audit trails.
Chain reconstruction: rebuild and export audit trails.
No parameter schema in public metadata yet.
swt3_trajectoryTrajectory attestation: prove autonomous decisions were recorded.
Trajectory attestation: prove autonomous decisions were recorded.
No parameter schema in public metadata yet.
Sponsors
Configuration
SWT3_API_KEYsecretSWT3 API key (starts with axm_). Omit for zero-config demo mode.
SWT3_TENANT_IDYour tenant identifier
SWT3_CONFIG_FILEPath to swt3.yaml config file for policy-as-code
SWT3_CLEARING_LEVELData clearing level: 0=analytics, 1=standard, 2=sensitive, 3=classified
SWT3_AGENT_IDAgent identity for AI-ID.1 witnessing
SWT3_SIGNING_KEYsecretHMAC-SHA256 signing key for non-repudiation
Readme
SWT3 - Sovereign Witness Protocol for AI
Don't audit the agent's thoughts. Audit the agent's actions.
The Problem
AI agents are making production decisions: approving loans, triaging patients, managing infrastructure, writing code. In 2026, 65% of firms reported AI agent security incidents. Only 14.4% of agents go live with full security approval. When something goes wrong, there is no tamper-proof audit trail. Logs are mutable. Metrics are averaged. Nobody can prove what the agent actually did.
Details
- Repository
- github.com/tenova-labs/swt3-ai
- Website
- tenova.io
- Package
@tenova/swt3-mcp- Transport
- STDIO
- Auth
- Required
- Registry
- MCP registry, active