INSTALL_FOR_AGENTS.md

Install Agent Vault MCP for an agent runtime

Use this when an agent runtime already has memory/execution but needs verifiable context capsules and owner-approved memory proposals.

Configure MCP

{
  "mcpServers": {
    "scopeblind-agent-vault": {
      "command": "npx",
      "args": ["-y", "@scopeblind/agent-vault-mcp"],
      "env": {
        "SCOPEBLIND_AGENT_MANIFEST": "/path/to/agent/manifest.json",
        "SCOPEBLIND_AGENT_PAGES": "/path/to/agent/signed-pages.json",
        "SCOPEBLIND_AGENT_VAULT_ALLOW_DEMO_AUTH": "false"
      }
    }
  }
}

For local development before npm publishing:

{
  "mcpServers": {
    "scopeblind-agent-vault": {
      "command": "node",
      "args": ["/path/to/scopeblind-agent-vault-mcp/src/index.mjs"],
      "env": {
        "SCOPEBLIND_AGENT_MANIFEST": "/path/to/agent/manifest.json",
        "SCOPEBLIND_AGENT_PAGES": "/path/to/agent/signed-pages.json"
      }
    }
  }
}

Agent protocol

Before a run:

1. Call vault_context_request with the page kinds you need.
2. Use only included_pages as model-visible context.
3. Treat redactions as policy constraints, not as missing data.

After a run:

1. Call vault_memory_propose with candidate durable learnings.
2. Do not claim the memory was saved until the owner approves and signs it.
3. If asked to prove what context was disclosed, call vault_capsule_verify.

One-sentence boundary

You may propose memory; you may not silently rewrite the user's Vault.