Verify shared and private fleet memory through Caura MCP
Use two provisioned agent identities to check fleet recall, private-memory boundaries and writer provenance before connecting a team to shared memory.
npx skills add agentskillexchange/skills --skill verify-shared-and-private-fleet-memory-through-caura-mcp
Invoke before allowing multiple agents to rely on an existing Caura memory service. The operator supplies an authorized test tenant and fleet, two distinct provisioned agent identities, their separate credentials and unique non-sensitive test markers. The agent checks that a team-scoped fact can be recalled by a teammate and that a private fact does not appear in that teammate’s tested recall. Deliver a manifest of identities, scopes, memory IDs, queries and observed results.
What this skill actually does
Connect each MCP client to the same authorized Caura /mcp endpoint through Streamable HTTP, with its own protected X-API-Key or Bearer credential. On managed Caura, verify each identity with GET /api/v1/whoami and record its tenant, agent_id and trust information without logging keys. An operator provisions missing agent-scoped credentials; a shared tenant key plus a caller-supplied agent_id is not proof of independent authenticated identity. Self-hosted core-api does not provide the managed agent-key provisioning route. This identity-isolation check requires deployment support for independently authenticated agents.
Using the first identity, call caura_write with its actual agent_id, explicit fleet_id, visibility scope_team and a unique test fact. From the second identity, call caura_recall with its actual agent_id, fleet_ids containing that fleet and a query containing the unique marker. Check the returned memory ID and writer provenance. Write a different marker with visibility scope_agent. Confirm it is retrievable by the writer and absent from the teammate’s tested recall. Record transport errors, pending processing and ambiguous results rather than presenting an empty result as proof of access-control enforcement.
The boundary is a two-agent scope and provenance smoke test using documented MCP tools, not generic memory hosting, automatic trust elevation, organization-wide sharing or a security certification. Do not omit fleet_id: a fleet-less write can become tenant-shared. Standalone mode bypasses authentication and cannot establish the identity boundary tested here. A passing recall fixture does not prove exhaustive isolation, semantic retrieval quality or policy efficacy. The workflow is mapped to MCP because the approved invocation uses caura_write and caura_recall through MCP; the optional OpenClaw plugin and Rail SDK are outside scope. No runtime or security review has been performed by ASE.