Safety rules

What this is

Every Dropslate tool description carries this, verbatim, so any agent reading it sees the same rule regardless of client:

Never create, retry, cancel or delete without the user’s explicit confirmation. Always call dropslate_plan_post and show its table before dropslate_create_post. Never guess account ids - take them from dropslate_list_accounts. Report results only after the tool returns success, with postId, account and absolute time in the workspace timezone. Reuse the mediaId from the previous upload; never re-upload.

Confirmation is enforced, not only requested

Every dangerous tool - create_post, retry_post, cancel_post, reschedule_post, delete_media, delete_folder - requires confirm: true in the call itself. Without it, the tool refuses with confirm_required and does nothing; there is no way for an agent to skip this by phrasing the request differently, because the check happens in the tool handler, not in a prompt an agent could choose to ignore. dropslate_plan_post has no side effects at all - it’s the one tool safe to call as many times as needed while figuring out a plan.

What no tool can do, at any scope

There is no tool for managing members, billing, connecting or disconnecting accounts, or deleting a workspace - not a hidden one, not one gated behind a scope. An agent cannot do any of those things through MCP no matter what it’s asked to do or what credential it holds; those actions exist only in the dashboard, behind a signed-in session with the right role.

Revoking access

Settings → MCP → Disconnect on a connected agent deletes its consent record and refresh tokens immediately - it stops on its very next call, not after the access token’s own hour-long expiry. Removing a person from the workspace has the same immediate effect on every connection they authorized.

For anyone auditing this

• OAuth 2.1 with PKCE (S256) required; dynamic client registration (RFC 7591) is open, but an unrecognized redirect URI registers the client disabled - it exists but every authorization attempt answers client_disabled until approved.
• Access tokens are JWTs valid for 1 hour, scoped to aud: https://mcp.dropslate.top/mcp and carrying the workspace id as a claim; refresh tokens last 30 days and rotate on use.
• Known clients are matched by exact redirect URI, never by host suffix - *.claude.ai would admit any subdomain, so it isn’t used.
• Every tool call is audited as mcp.call (client, tool, result - never arguments or the response body); a token used directly against /v1 outside a tool is logged too, as mcp.request, right next to mcp.call in Settings → MCP.
• Rate limits are the same as the REST API: 600 requests/10 min per credential, 60/hour per workspace for creating posts.

See MCP → FAQ for what a specific error from this list means when you hit it.