Security & Approval
Tecrübe's security model: OAuth 2.1, preview → approval, budget guardrail, audit log, real undo and results the assistant cannot invent.
Three principles
- Everything that changes your account is previewed first. Write tools do not go straight to the platform; the action engine builds the plan and holds it.
- Approval comes from you. The assistant may call
approve_actiononly after you explicitly approved, withconfirm=true. Unless your approval policy is "automatic", nothing is applied on its own. - The result tells the truth. Every write tool returns applied, pending approval, dry run (no connection) or failed + reason. The assistant must have seen this result before saying "applied".
Identity and authorisation
- The assistant connection uses OAuth 2.1 + PKCE; there are no API keys. Clients register dynamically (DCR) and receive only the
mcpscope. - Platform connections (Google, Meta, LinkedIn, TikTok, Microsoft) go through the provider's own OAuth screen; Tecrübe never sees your password. The ChatGPT Ads key and the WordPress application password are stored encrypted in the panel.
- Tokens are specific to a member and an assistant and are revoked individually from the panel. A call with a revoked token is rejected.
- In team accounts every member connects with their own identity; the audit log records who called what.
Budget guardrail
Every budget change passes two limits; a request beyond them is not applied even if approved, and says why:
| Limit | Rule | Example |
|---|---|---|
| Single step | A budget rises by at most 50% per change | 100 → at most 150 |
| 24 hours | The total increase in one day cannot exceed twice the starting budget | A campaign starting the day at 100 cannot pass 200 that day |
Undo is not caught by this limit: undo_action returns to the previous budget. In addition, the spend guard (set_spend_guard) lets you set daily/monthly caps, target CPA and margin; automation rules stay within those caps.
Audit log
Every tool call is recorded: which member, which assistant, which tool, parameters, duration, credits spent, result and error. Write operations additionally keep the preview, approval time and application result in the action log. Records are visible in the panel under Usage and Actions; the list_recent_actions tool brings the same log into chat.
Real undo
"Undo" does not just delete the record in Tecrübe; a real reverse request goes to the platform: back to the previous budget, resume after pause, remove added negative keywords, unpublish, restore a WordPress revision, revert a GTM version. Operations with no reverse (a sent message, a removed campaign) return not_undoable with the reason; the tool says so instead of erroring.
Data and privacy
- Tecrübe passes tool results to the assistant; it does not store your chat content. The assistant provider's (Anthropic, OpenAI…) own data policy applies.
- Platform tokens are stored encrypted and decrypted on the server only during a tool call.
- Details: Privacy policy and data protection notice.