Register an MCP tool server once, then route every tool call through
POST /v1/mcp/execute. IdentArk applies policy, pauses high-risk calls for human
approval (HITL), and writes an audit entry for each decision.
Risk is computed, not hardcoded
Risk scores are derived from the properties of each call — operation class, data
sensitivity, scope breadth, environment, and taint — not fixed per tool. The same
tool scores higher when its scope widens: an erase with an unbounded where clause
scores far above a targeted one, which is how an injected instruction gets caught and
refused.
Endpoints
| Method | Path | Scope | Purpose |
|---|---|---|---|
POST |
/v1/mcp/servers |
org:admin |
Register a tool server. |
POST |
/v1/mcp/servers/{id}/discover |
org:admin |
Enumerate its tools. |
POST |
/v1/mcp/execute |
key-scoped | Run a tool call under policy + audit. |
POST |
/v1/mcp/policies |
policies:write |
Define HITL / approval thresholds. |
GET |
/v1/mcp/approvals/pending |
policies:read |
Approvals awaiting a human. |
POST |
/v1/mcp/approvals/{id}/decision |
policies:write |
Approve / deny a paused call. |
POST |
/v1/mcp/kill-switch |
org:admin |
Halt all execution instantly. |
GET |
/v1/mcp/audit/chain/verify |
audit:read |
Verify audit-chain integrity. |