AI connection: what to do when it fails

“Unauthorized” or the client asks to reconnect

The access token lasts 1 hour and the refresh token 60 days. Clients refresh silently; if the refresh token has expired, or you removed the connector in Assetli, the client has to go through consent again. That is expected. If it happens every few minutes, the client is not storing the refresh token; reconnect from scratch by removing the connector on both sides.

The consent screen appears again on every connection

Some clients register a new OAuth client each time they connect (dynamic client registration). Each registration is a new connector in Assetli’s eyes, so consent is asked again. It is harmless; old connectors without tokens expire on their own. Remove leftovers under the account menu → AI Connection (MCP).

Write tools are not offered to the AI

The tool list a client receives depends on two things: the Write scope you granted on the consent screen, and a Premium or Max plan (the 30-day trial counts). If either is missing, only read tools are listed. After the trial ends, write tools disappear from the list without any change on your side; after an upgrade, they appear on the next connection. Clients cache the tool list, so disconnect and reconnect to refresh it.

”This operation requires the write scope”

You granted read-only access. Remove the connector in Assetli and connect again, choosing Read and write on the consent screen.

”Writing and importing over MCP are part of the Premium plan”

Reading is free on every plan; writing needs Premium or Max. Upgrade under the account menu → My plan.

”Missing confirmation_token”, “expired”, “arguments differ from the preview”

Every write requires a token from its preview tool, valid for 10 minutes, bound to the exact values previewed. The messages tell the model what to do: run the preview again, show it to you, and pass the new token. If the model keeps failing here, ask it explicitly to call the preview tool first. This is the server refusing a write that was not previewed; it is not a bug in your connection.

The AI sees the wrong profile, or not all profiles

A connector is bound to the one profile you chose on the consent screen and cannot switch. To use another profile, remove the connector and connect again, choosing the other profile. To use several profiles at once, connect several times and name them differently in the client.

ChatGPT reads but refuses to write

At the time of writing OpenAI allows write-capable custom connectors only on Business, Enterprise and Edu workspaces. On Plus and Pro the write scope you granted in Assetli is not used. This is OpenAI’s restriction; check their documentation for the current state.

Gemini cannot add the connector

The consumer Gemini app does not support custom MCP connectors at the time of writing. Gemini CLI and Gemini Enterprise can add the server with the Streamable HTTP transport and OAuth.

”Unknown tool”

The client is using a cached tool list that no longer matches the server, for example after a tool was renamed. Disconnect and reconnect to refresh the list.

The connector URL

https://assetli.app/api/mcp, with the Streamable HTTP transport. The server answers JSON-RPC over POST only; it does not offer a server-initiated stream (GET returns 405), and it offers tools only, not resources or prompts.

Checking what happened

Under the account menu → AI Connection (MCP) you see every connector with its scope and profile, and the history of writes made through AI: tool, connector, time, result and a summary of the arguments. Reads are not logged.