Agents that work within your rules.#
- Scoped tool calls. Every MCP tool call carries the scopes of the API key or session it was launched under. A tool can never escalate beyond what the caller is already allowed to do.
- RBAC, not bypass. An agent calling create_post obeys the same approval rules a human would: drafts land in the right approval state, never auto-publishing past required reviewers.
- Destructive actions require confirmation. Deletes, role changes, and bulk operations require an explicit confirm token. Agents can plan and propose, but a human or signed automation has to confirm the irreversible step.
- Every call is audited. Agent tool calls produce the same audit events human actions do — actor, scope, target, payload, timestamp. Agent activity is visible in the same audit log as human activity.
- Provider capability awareness. Tools like create_post run the same capability validation a human's calendar entry would. An agent can't schedule a post a platform won't accept.
- Memory respects workspace boundaries. Brand voice memory and campaign memory used by agents are workspace-scoped. Nothing leaks across workspaces, even when the same agent is connected to multiple.
What ships in the MCP server today.#
- list_workspaces. Enumerate the workspaces this caller can see, with role and scope per workspace.
- create_post. Draft a post into a workspace's calendar with capability validation and approval routing.
- upload_media_from_url. Pull a remote asset into the workspace media library with scan, processing, and licensing metadata.
- get_calendar. Read scheduled and shipped posts with status, channel, and approval state.
- get_pending_approvals. List approval items waiting on the caller or a named approver.
- draft_reply. Draft a reply for an inbox thread — never sends; queues for human acceptance.
- get_analytics. Pull metric rollups with source and freshness metadata preserved.
- generate_client_report. Produce a report draft with cited metrics, disclosed gaps, and white-label-ready output.
Agents should work, not surprise you.#
An agent that bypasses your approval rules is an incident waiting to happen. SignalDeck's MCP server runs agents inside the same accountability surface as humans — same audit, same RBAC, same confirmation gates on destructive work.
Questions#
- Which MCP clients work?
- Any compliant MCP client. The server speaks standard MCP transports and ships an example wiring for Claude Desktop, Cursor, and direct HTTP.
- Can an agent publish without a human?
- Only if the workspace policy allows it for the role/scope the agent runs under. Default policy requires human acceptance for sensitive actions and publication.
- How do I see what an agent did?
- Every tool call appears in the workspace audit log alongside human actions. Filter by actor type to see only agent activity.