Sanctuary on Claude Managed Agents

Add sovereign identity, encrypted audit trails, and verifiable reputation to your Managed Agent in under 60 seconds.

Prerequisites

Step 1: Add Sanctuary to Your Agent Config

In your Managed Agent definition, add Sanctuary as an MCP server:

agent:
  model: claude-sonnet-4-5-20250514
  system_prompt: |
    You have Sanctuary Framework available via MCP.
    On startup, call sovereignty_health_report to verify your status.
  tools:
    - type: agent_toolset_20260401
  mcp_servers:
    - name: sanctuary
      command: npx
      args: ["@sanctuary-framework/mcp-server@1.8.4"]

Deploy the agent. Sanctuary tools load automatically via MCP.

Step 2: Verify Sanctuary Is Active

Call manifest from your agent. You should see the Sanctuary tools listed (more than 100 in v1.8.4).

Step 3: Bind an Identity

Call: identity_create
Call: identity_set_primary with the new identity ID

The agent receives a DID like did:key:z6Mk... backed by an Ed25519 keypair.

Step 4: Run a Sovereignty Health Report

Call: sovereignty_health_report

See the verification checklist for expected results.

Step 5: Publish to Verascore (Optional)

Call: reputation_publish

From 1.9.0 this call always waits for operator approval (Tier 1), because the signed report leaves the machine. Releases before 1.9.0 publish without asking under the default policy.

Visit https://verascore.ai/agent/{did} to see the live profile.

What You Get

Before Sanctuary After Sanctuary
Implicit cloud identity Explicit agent-owned Ed25519 keypair
Default logging Encrypted, tamper-evident hash-chained audit trail; production audit checkpoints are currently unsigned until IC-05 closes
Platform-level policy (static) Principal-level policy (dynamic, time-locked)
All-or-nothing disclosure Selective disclosure (ZK attestations)
Opaque reputation Verifiable Verascore profile

Managed Agents-Specific Notes

Container environment. Node.js is pre-installed in Managed Agents containers. npx works out of the box.

Network access. Sanctuary’s Verascore publishing (reputation_publish) requires outbound HTTPS to verascore.ai. Ensure your environment’s network rules allow it.

Identity persistence. Container filesystems persist within a session but not across sessions. For durable identity across sessions:

Multi-agent. If using callable_agents, each thread agent can independently run Sanctuary: one identity per thread agent, or a shared identity via seed.

SIEM export. Use audit_export_siem to export CEF, OCSF, ECS, or raw Cortex events for XDM modeling from inside the container. Field names under integrations/siem/, and the Cortex mapping file’s raw-to-home pairs, are checked against formatter tables by repo tests; null XDM homes keep those values as raw dataset fields. Run a tenant drill before treating either mapping as installed evidence.

Troubleshooting

“npx: command not found” Unlikely in Managed Agents (Node.js is pre-installed). If using a custom environment, ensure Node.js 18+ is in PATH.

Verascore publish fails with network error Check environment network access rules. verascore.ai must be in the outbound allowlist.

Identity lost between sessions Expected behavior: container filesystems are ephemeral. Export the identity keypair at session end, or use a deterministic seed for consistent DID generation.

Tool confirmation prompts for Sanctuary tools If tools are configured with confirmation requirements, Sanctuary’s audit and identity tools will trigger them. Consider whitelisting Sanctuary tools in your tool confirmation config.

First run is slow npx downloads the package on first invocation (~10s). For faster cold starts, add RUN npm install -g @sanctuary-framework/mcp-server@1.8.4 to your container setup.

Next Steps