Troubleshooting

MCP Connected But Still Not Working?

If your connector shows as installed but your agent still says it can't access the tool, the issue is usually auth/session state — not the model.

This is one of the most common setup traps: you installed a connector, but calls still fail with "not available," "not authorized," or "please install MCP first." Usually, one layer is still stale or incomplete.

Important: "Installed" does not always mean "authorized and active in this session." Connector install, OAuth auth, and session context are separate steps.

Fast recovery checklist (5 minutes)

Confirm the connector appears installed and enabled.
Re-open the connector and confirm OAuth/account auth is fully complete.
Start a fresh chat session after auth (old sessions can keep stale state).
Run a minimal smoke test prompt (example below).
If it fails, disconnect/reconnect once, then retest in a fresh session.

Known-good smoke test prompt

Use the [TOOL NAME] connector and do the smallest safe read-only check. Return: 1) whether the connector is available, 2) whether auth is valid, 3) the exact error text if it fails. Do not perform any write actions.

Most common root causes

OAuth flow started but not fully completed (tab closed early).
Connector installed in one workspace/account, tested from another.
Old chat session still holding pre-install context.
Permissions mismatch (connector has limited scope and can’t access target resource).
Provider-side token expired and needs reconnect.

Quick isolation test

Test with a read-only request on a known resource you definitely own.
If possible, test a second resource in the same account.
If one works and one fails, it's likely a resource permission issue, not connector install.

Avoid this trap: Don’t keep retrying in the same long chat. A fresh session after reconnect fixes a surprising number of “still unavailable” loops.

“It keeps forgetting I installed it” — what’s actually happening

Users often describe this as the connector being "forgotten," but it’s usually one of these context mismatches:

Different surface, different session: You installed in dashboard chat, then tested in Discord (or the reverse) without a clean handoff.
Different account/workspace: You authenticated the connector under one account, but your active agent/chat is in another.
Session pre-dates install: The chat was opened before install/auth completed and still has stale tool state.

Fast fix: After install + OAuth, start a brand-new session in the same surface you plan to use, then run the read-only smoke test once before doing real work.

Support packet (copy/paste)

Connector issue report: - Connector name: [e.g., Dropbox MCP] - Installed + enabled: yes/no - OAuth reconnected today: yes/no - Fresh session tested after reconnect: yes/no - Read-only smoke test prompt used: [paste] - Exact error text: [paste full text] - Workspace/account where connector is installed: [name] - Resource tested: [what you tried to access]