skovos

Skovos Governance connector: documentation

Skovos Governance gives health systems a registry of record for their AI agents, explicit permission checks, a tamper-evident audit trail, a stop control the hospital owns, and signed evidence that those controls ran. This connector exposes those capabilities to Claude and ChatGPT over the Model Context Protocol (MCP).

Connect

Registering grants, expanding scopes and recalling agents require the separate operator API. MCP clients cannot approve themselves or recall other agents. Workload tools require a unique request_id for retry safety. Organization overview, registry reads, audit reads and evidence exports are available to observers only.

Vendor reporting skill

Both connector routes advertise skovos://skills/governance, the current vendor reporting skill from vendor.skovos.ai. Server instructions point to it when a client connects. Clients without resource support can use get_governance_skill. Automatic loading depends on the client.

The skill is metadata-only guidance. Reading it does not authorize or submit reports, connect a vendor account, or enable patient-data submission. Vendor reporting requires separate operator authorization and vendor API credentials. Never send patient data.

Tools

ToolWhat it doesChanges data?
get_governance_skillFetch the public vendor reporting skill from its fixed source URL; no caller data or credentials are forwardedNo
governance_overviewOne-screen summary: agents, owners, scopes, denials, stops, audit integrityNo
list_agentsEvery registered agent with status, owner and scopesNo
check_permission (workload code only)Allow or deny an action for an agent; the decision is auditedAdds an audit entry
record_action (workload code only; self-reported)Append a hash-chained audit record of an agent action. Text matching common identifier patterns (SSN, MRN, email, phone, dates) is rejected. Names and addresses are not detected, so do not send patient data. This heuristic screen runs on record_action only.Adds an audit entry
get_audit_trailRecent audit entries, filterable by agent or resourceNo
verify_audit_integrityRecompute the hash chain and report tamperingNo
scan_phiHeuristic PHI detector with a redacted copy (Claude endpoint only; not a certified de-identification tool)No
export_overt_evidenceSigned, hash-chained evidence records (OVERT v1.1 style, self-declared profile, not a conformance claim). No agent names or content in the bundle.Only on first use: creates the organization's signing key
verify_overt_evidenceVerify an evidence bundle's signatures, chain and countersNo

Example prompts

Limits

Permission checks are decision receipts, not reusable execution permits. Action reports are labeled self-reported. Recall stops subsequent checks through this connector; native vendor actions require a separately integrated execution boundary. It does not observe systems it is not connected to. Evidence export uses a self-declared profile pending OVERT registration.