Saltar para o conteúdo
Tecrübetecrube.com
First steps

How it works

The journey of a chat message through Tecrube to the ad platform and safely back: MCP, OAuth, guardrails, approval of suggestions, audit log and undo.

The journey of a request

  1. You type. You tell your assistant (Claude, ChatGPT, Cursor…) an ordinary sentence: "Add the search terms with no conversions as negatives."
  2. The assistant picks a tool. Over MCP it sees Tecrube's tool list and calls the right one (e.g. search_terms_report, then add_negative_keywords) with parameters.
  3. Tecrube authenticates. Every call carries the access token you granted through OAuth 2.1; the token says which member and which account it belongs to.
  4. Read tools answer directly. Data is pulled from the connected platform; if no account is connected, the tool says "not connected" instead of inventing numbers.
  5. The change you ask for is applied at once. The write tool goes through the action engine: what changes, in which account, and the old and new value are recorded. The budget guardrail and spend cap step in here; a request over the limit is not applied and says why.
  6. The assistant's own suggestions wait for you. Changes you did not ask for (automation rules in "ask" mode, proactive suggestions) land in the panel's approval list and reach the platform only after you approve them. Permanent deletions are confirmed with you in chat first.
  7. It is logged and reversible. Every applied change is written to the audit log, and reversible operations can be undone with undo_action.

What is MCP, and why hosted?

The Model Context Protocol (MCP) is an open protocol that lets AI applications discover and call tools in external systems in a standard way. Tecrube serves this protocol over the Streamable HTTP transport at a single address hosted on our infrastructure:

MCP URL
https://tecrube.com/mcp

You do not install anything on your computer or run Node or Python. Desktop, web and mobile clients connect to the same address; the tool list is updated on the server and nothing changes on your side.

Authentication: OAuth 2.1 + PKCE

There are no API keys. On first connection the client reads the .well-known/oauth-protected-resource document, finds Tecrube's authorization server, registers itself dynamically (DCR) and sends you to the Tecrube sign-in page in your browser. After signing in you see which assistant is connecting on the consent screen and approve it; the client exchanges a PKCE-protected authorization code for an access token.

  • Tokens are specific to a member and an assistant; each can be revoked from the panel.
  • When the access token expires the refresh token is used; you are not asked to sign in again.
  • The discovery document is public and contains no data:
Discovery document
https://tecrube.com/.well-known/oauth-protected-resource

Safety layers

Requests and approvals

Changes you ask for are applied at once and can be undone; unprompted suggestions and rules in "ask" mode wait for your approval.

Budget guardrail

A budget rises by at most 50% per change; the total increase within 24 hours cannot exceed twice the starting budget. Requests beyond that are not applied even if approved.

Audit log

Every tool call is recorded: who, when, with which parameters and what the result was. Visible in the panel.

Real undo

Budget, pause, negative keywords, publishing, WordPress edits and similar operations are reversed with a real request to the platform.

See Security & Approval for details.

Open problems and the nightly scan

  • The assistant scans your connected accounts every night; it records problems it finds in delivery, policy, budget, bidding, tracking, connection, sync, product feed, SEO, production and plan, and closes them once resolved.
  • The Open problems card on the Overview screen lists them by severity; Solve with the assistant opens the chat with a ready request, and you send the message.
  • On its own the scan does only safe maintenance (syncing, renewing authorisations, re-queueing a stuck free production); it does not change budgets, bids or publishing. Such fixes are suggested to you.
  • In MCP clients, list_open_problems returns the same list and each problem's possible fixes, read only.

Next steps