MCP can now write, not just read — The PartnerBoost MCP server adds 13 write tools alongside the existing read tools, so your AI assistant can carry out actions instead of only answering questions. The MCP URL and AI Access Key do not change — reconnect your client to pick up the new list.
Per-key scopes — Each AI Access Key now carries its own scope. Only the capabilities you grant appear in that key's tools/list, and calls outside the scope are rejected. Write capabilities are grouped into four packs:
Partners — Manage partner workspaces, decide on relationship changes, invite or message partners.
Catalog & commission — Manage products, creatives, commission rules, Amazon promos, coupon tracking, and ads budget. Amazon promo tools appear only for Amazon brands.
Outreach — Manage newsletter templates and groups, send newsletters, and configure recruitment and PPC.
Orders — Create manual affiliate orders and resolve order inquiries.
Preview before you commit — Write tools accept dry_run, which validates the call and returns what would change without persisting anything.
Idempotent writes — Every write call requires an idempotency_key. Replaying the same key on the same tool does not write twice; reusing a key with different content returns a conflict error.
Write audit trail — Every write call is logged, including denied ones, so actions can be traced after the fact.
Manage multiple AI Keys — Generate, rename, and archive multiple AI Keys per account (up to 10), each with its own scopes. Sub-accounts see a contact-your-admin hint and cannot generate or manage keys themselves.
Stricter API Key permissions — API Keys are now checked against the permissions of the account that owns the key. A key created by the main account keeps full access; a key created by a sub-account only gets the permissions granted to that sub-account, and can no longer fall back on the main account's access. If a sub-account key now returns Permission Not Found, grant that sub-account the matching permissions.
Clearer errors — Unsigned agreement, invalid key, missing scope, missing backend permission, and validation failures now return consistent, readable error codes so the assistant can decide what to do next.