SWT3 Witness Connector for Meta Muse

1. What SWT3 Does in Muse

SWT3 is a passive compliance witness. It observes your Muse agent's actions and creates tamper-proof cryptographic evidence of what happened. It never modifies agent behavior, never adds latency to the critical path, and never sends your prompts or responses to any server.

When connected to Muse, SWT3 records evidence for:

• Every model inference your agent makes
• Every tool call and its outcome
• Authorization decisions and permission checks
• Agent-to-agent handoffs and delegation chains
• Human review checkpoints
• Consent and lawful basis documentation
• Service account credential usage

Each event produces a witness anchor with a unique cryptographic fingerprint that anyone can independently verify has not been altered.

2. One-Message Setup (Stdio)

Muse natively supports MCP servers. The fastest integration path is a single message in any Muse conversation. Copy and paste:

Build a custom integration to the SWT3 AI Witness MCP server.
Install it with: npx @tenova/swt3-mcp
It runs over stdio with Node.js. No authentication is needed for
demo mode. Connect to it, test the witness_inference and
list_procedures tools end to end, show me the results, and save
the integration as a reusable skill.

This installs the MCP server on Muse's VM and connects immediately. Demo mode creates local anchors with no account required.

Connected mode (persistent ledger)

To persist anchors to a cryptographic ledger where auditors can verify them, add your API key:

Build a custom integration to the SWT3 AI Witness MCP server.
Install it with: npx @tenova/swt3-mcp
It runs over stdio with Node.js. Set the environment variable
SWT3_API_KEY to my key (I'll provide it through the secure
credential flow). Connect, test witness_inference and
list_procedures, and save as a reusable skill.

Muse stores credentials through its Secure Credentials Store, separate from the agent runtime. Your API key is never visible in conversation text.

No API key yet? Use the signup tool inside Muse to create a free account and get one instantly.

3. URL-Based Setup (Streamable HTTP)

For teams running a self-hosted SWT3 MCP endpoint, Muse can connect via URL instead of installing the npm package.

Build a custom integration to the SWT3 AI Witness MCP server.
It is a remote MCP server over streamable HTTP at:
https://your-endpoint.example.com/mcp
No authentication is needed for the MCP connection itself.
Connect to it with the official MCP SDK, test witness_inference
and list_procedures end to end, and save the integration as a
reusable skill.

Running the HTTP endpoint

# Start the streamable HTTP server (localhost:3100)
npx @tenova/swt3-mcp-http

# With API key for persistent anchors
SWT3_API_KEY=axm_live_... npx @tenova/swt3-mcp-http

# Custom port
SWT3_MCP_PORT=3200 npx @tenova/swt3-mcp-http

The server binds to 127.0.0.1 by default. Place a reverse proxy (Caddy, nginx) in front for TLS and public access.

4. What Gets Witnessed

SWT3 exposes 68 witnessing tools via MCP. The ten most relevant for Muse agent governance:

All 68 tools are available to Muse. Run list_procedures in any conversation to browse the complete registry filtered by framework, namespace, or keyword.

5. Clearing Levels

You control exactly what information leaves the machine. SWT3 clearing levels determine data retention granularity:

At every level, raw prompt and response text stays on your machine. Only cryptographic hashes and numeric factors are transmitted.

Set the clearing level via environment variable:

SWT3_CLEARING_LEVEL=2 npx @tenova/swt3-mcp

6. Demo vs. Connected Mode

Demo mode is fully functional for evaluation. Every tool works identically. The only difference is where anchors are stored.

To upgrade from demo to connected mode, use the signup tool inside Muse:

Use the signup tool to create a free SWT3 account.

This creates your account and returns an API key without leaving the conversation.

7. Framework Coverage

SWT3 witnesses map to 77 regulatory frameworks through the UCT (Unified Control Taxonomy) crosswalk engine. Frameworks most relevant to Muse deployments:

Use resolve_crosswalk inside Muse to look up how any procedure maps to your specific framework.

8. How It Works

The response is already on the wire before the witness fires. SWT3 cannot add latency, cannot cause tool failures, and cannot modify agent behavior. This is a deliberate architectural decision: compliance evidence collection must never interfere with the system it observes.

9. Common Questions