AI · MCP server
The agent's toolset, in whatever MCP client your team already uses.
Cursor, VS Code, Claude Code, Claude Desktop — CloudPivot's built-in MCP server exposes the same tool registry as the in-app agent, scoped by workspace key. Read-only and draft-only, always.
One registry, zero drift
Same tools, same policy, same audit trail as the in-app agent.
There is deliberately no second tool catalog. Every MCP call routes through the exact registry the built-in chat uses, so disabled-tool policy, budgets, timeouts, and redaction apply identically — and every call lands in the same audit log, attributed to the key that made it. Admins pick which tools each workspace exposes; switch one off and it disappears from every editor at once.
Keys are the tenant boundary
Workspace-scoped cpk_ keys are the only credential the endpoint accepts — stored as SHA-256 hashes, shown once at creation, created and revoked by workspace admins, and valid nowhere else in the API.
Fail-closed gates before any tool runs
Valid key, workspace MCP switch, entitlement, per-key rate limit — in that order. Any failure returns before tool dispatch, and tools/call re-checks exposure at execution.
Read-only and draft-only, enforced
Tools that apply changes stay behind the in-chat human-approval flow and are hard-excluded from MCP. Nothing reachable over MCP mutates a cloud resource.
Modern protocol, older clients welcome
Stateless Streamable HTTP per the MCP 2026-07-28 revision — no sessions, no handshake — with a legacy shim so pre-2026 Cursor and Copilot builds connect too.
The trust path
One more front door to the same audited tool loop.
An MCP tool call writes the same audit rows as an in-chat call — same tenant scope, same redaction — so security reviews one surface, not two.
Bring CloudPivot to your editor.
Enable MCP in workspace settings, create a key, and paste the generated client config — setup is a copy-paste, not a deployment.