2.5 · Fix a broken MCP connection

Fixes for the moving parts: an expired sign-in, an unloaded tool registry, and naming mismatches.

MCP connections are stable once working, but they do have moving parts: an OAuth sign-in that expires, a tool registry your agent has to load, and names that have to resolve to the right workspace (the container holding everything GrowthOS knows and does for your brand) and page. When your agent stops answering GrowthOS questions, it's almost always one of the 4 cases below.

Expected behavior, not a bug. Your sign-in uses short-lived credentials that refresh automatically while you're active, but after roughly 30 days without use the refresh itself expires, and errors like "Refresh token expired" mean it's time to reconnect. Re-run the authenticate step for your tool: reconnect the connector in Claude, /mcp then Authenticate in Claude Code, or Connect again in Cursor. Modern MCP clients usually detect the expiry and walk you straight into the sign-in.

The agent says it can't find GrowthOS tools

The server isn't registered or the session predates it.

Confirm the server is added in your tool's MCP settings (the server name is growthx-os) and the URL matches the one on your MCP Settings page.
Start a fresh session or conversation; tools load at session start in most agents.
In Claude Code, /mcp shows the server and its status; if it's listed but unauthenticated, that's the cause.

If the server was never added at all, run setup from the top.

The agent answers, but about the wrong things

Wrong workspace or a name that doesn't resolve.

If you have access to several workspaces, say which one you mean, by name. The agent can list the workspaces it sees; if one you expect is missing, your GrowthOS account doesn't have a seat there.
"Page not found" style errors usually mean the page was referenced by an approximate name. Give the URL path (like /blog/pricing-guide) or ask the agent to search for the page by title first, then work from what it finds.

The connection was denied or never completed

A consent flow that ended at Deny, or closed mid-flow, leaves the tool half-configured. Remove the GrowthOS server from your tool's MCP settings, add it again, and complete the sign-in through to the approval screen. The consent screen only ever asks for read access; there's nothing to configure on it beyond approving.

Still stuck: the 2-minute reset

The universal fix, in order:

Remove the GrowthOS server or connector from your tool.

Re-add it with the URL from your MCP Settings page.

Sign in when prompted, through to the approval screen, then start a fresh session and ask a workspace question.

Because credentials are per person, one teammate's broken connection says nothing about anyone else's, and resetting yours affects only you.

Common questions

The MCP Settings page says Connected, but my agent can't reach it. The badge records that a connection has succeeded from your account; it isn't a live health check. Trust the agent's behavior over the badge and run the reset above.

Is publishing broken if MCP is broken? The agent publishing path needs two things: this connection (to read the brief, the plan for one page) and your CMS's own MCP (to write the draft). If reading works but drafting fails, the thing to troubleshoot is the CMS integration on your agent, not GrowthOS.