If your MCP server is not connecting or your MCP tools are not showing up in Claude, the cause is almost always one of a handful of predictable issues: an expired authorization token, a wrong endpoint URL, a connector that got disabled mid-conversation, or a tool list that hasn’t refreshed after a server update. This guide walks through each one in order, from the most common to the most easily overlooked.
The fixes below apply to any MCP server connected to Claude or ChatGPT, with Adsroid MCP used as a worked example throughout since it covers a range of real-world connection scenarios that most troubleshooting guides skip.
What Is an MCP Server and Why Does Connection Matter?
The Model Context Protocol (MCP) is an open standard that lets AI assistants like Claude connect to external tools, data sources, and services through a structured server interface. When working correctly, the AI assistant discovers available tools from the MCP server at the start of a session, calls them during the conversation, and returns results inline.
When the connection breaks, the assistant either loses access to all tools silently, shows a reduced or empty tool list, or returns errors when attempting to call a tool. The frustrating part is that the AI often continues to respond as if everything is fine, just without any real data behind its answers.
An MCP connection failure does not always produce an obvious error message. The assistant may simply stop calling tools and fall back to generic responses. If your AI assistant’s answers suddenly feel less specific or data-driven, check the connector status before anything else.
Fix 1: Re-Authorize the MCP Connector (Expired Token)
This is the most common cause of a sudden MCP server not connecting. OAuth access tokens expire. Depending on how the MCP server is configured, tokens can expire anywhere from one hour to several days after the last authorization.
In Claude.ai, go to Settings > Connectors, find the relevant MCP server, and check whether it shows an authorization warning or a disconnected status. If it does, click to re-authorize and complete the OAuth flow again.
With Adsroid MCP, authentication is handled through an Adsroid API key rather than a short-lived OAuth token, which makes expiry less frequent. If you do see an MCP authentication failed error, the fix is to open the connector settings in Claude.ai, re-enter your Adsroid API key, and save. The connection restores immediately without needing to reconfigure anything else.
Fix 2: Verify the Endpoint URL Is Correct
MCP servers are identified by a URL. If that URL is wrong by even a single character, the connector cannot reach the server at all. This often happens after copying a URL from a browser that added tracking parameters, or after a server migration that changed the endpoint path.
Check that the URL in your connector settings matches the one documented by the MCP server provider exactly. For Adsroid MCP, the correct endpoint is mcp.adsroid.com/mcp. No trailing slash variations, no additional path segments. You can verify the current endpoint in the Adsroid MCP documentation.
If you are connecting to a self-hosted MCP server, also confirm the server is reachable from the internet and that no firewall rule is blocking inbound requests from Claude’s or ChatGPT’s IP ranges.
Fix 3: Check Whether the Connector Is Enabled in the Current Conversation
Claude.ai allows connectors to be toggled on or off at the conversation level. A connector can be fully authorized and working but disabled for a specific chat. This is a common source of confusion because the tool list appears empty even though the connection itself is healthy.
In Claude.ai, look for the connector controls at the bottom of the message input area. If you see the MCP server listed but toggled off, enable it. The tool list should populate within a few seconds. If you are starting a new conversation, the connector may default to disabled depending on your workspace settings.
ChatGPT’s connector behavior varies by implementation, but the same principle applies: check whether the integration is active for the current session, not just at the account level.
Fix 4: Force a Tool List Refresh After a Server Update
When an MCP server adds, removes, or renames tools, the AI assistant does not always pick up the changes immediately. The tool list is typically fetched at the start of a session and cached for the duration of the conversation. This means your assistant may be working from a stale tool list that doesn’t reflect the current server state.
The fix is straightforward: start a new conversation. This forces the client to fetch a fresh tool list from the server. If the new tools still don’t appear, disconnect and reconnect the connector in settings, then start another new conversation.
This situation comes up after platform updates. Adsroid MCP, for example, covers 140+ tools across Google Ads, Meta Ads, Google Analytics 4, Google Search Console, and Ad Radar. When new tools are added in a release, existing sessions will not see them until a new conversation is started.
Fix 5: Resolve OAuth Scope Errors
An MCP OAuth error that isn’t simply an expired token often points to a scope mismatch. This happens when the authorization was completed but the requested permissions were not fully granted, either because the user declined some permissions during the OAuth flow or because the account does not have the necessary access level.
To fix this, disconnect the connector entirely, then reconnect it and carefully review every permission request during the OAuth flow. Approving all requested scopes is usually required for the tool list to populate correctly. If you are authorizing on behalf of a client account, confirm you have admin-level access to that account before attempting the OAuth flow.
Fix 6: Confirm the Ad Account or Data Source Is Linked
Some MCP tools only appear or function once a specific data source is connected to your account. If your Google Ads or Meta Ads account isn’t linked to the platform backing the MCP server, the tools that depend on that account will either not appear or will return empty results when called.
This is especially relevant for advertising-focused MCP servers. With Adsroid MCP, the AI assistant gets access to whatever accounts are already connected within your Adsroid project. If you haven’t connected a Google Ads account to Adsroid yet, the Google Ads tools are present in the tool list but cannot return data. The fix is to connect the account through the Adsroid web app first, then return to Claude and try again.
If you are managing multiple clients, make sure the correct project is selected before running any tool call. Each project in Adsroid is fully isolated, so a tool call in the wrong project returns data for the wrong client, or nothing at all if that project has no accounts connected.
Fix 7: Check Plan or Workspace Restrictions
MCP connector access is not always available on every plan tier. Claude.ai, for example, gates certain connector features to Pro or Team plans. If the connector option is not visible in your settings, or if you can add the connector but it immediately shows an access error, the issue may be your subscription level rather than the connector configuration.
Similarly, some workspace administrators restrict which connectors team members can add. If you are on a Claude Team or Enterprise plan managed by your organization, ask your workspace admin whether external MCP connectors are permitted.
On the MCP server side, check whether your subscription covers API access. Some platforms offer MCP access only on higher-tier plans.
Fix 8: Clear Browser Cache and Retry
This sounds basic, but a corrupted session cookie or stale cached response from the AI client’s web interface can cause connector states to display incorrectly. Claude.ai is a web application, and like any web app it can serve stale UI state from cache.
Try a hard refresh (Ctrl+Shift+R on Windows, Cmd+Shift+R on Mac), or open Claude in a private browsing window and check whether the connector status looks correct there. If the connector works in a private window but not your main browser, clearing site data for claude.ai resolves it.
Fix 9: Test the MCP Server Endpoint Directly
If you have exhausted the above steps and the connection still fails, verify that the MCP server itself is responding. You can do this by sending a basic request to the endpoint URL using a tool like curl or Postman, or simply pasting the endpoint URL into a browser to see whether you get a recognizable response rather than a timeout or 404.
A timeout usually indicates a network or firewall issue between the AI client and the server. A 401 response points to an authentication problem. A 404 means the endpoint path is wrong. Each of these responses maps directly back to one of the earlier fixes in this guide.
For hosted MCP servers like Adsroid MCP, you can also check the provider’s status page or support documentation to rule out a service outage before spending time debugging your own configuration.
Fix 10: Reconnect the MCP Server from Scratch
If nothing else has worked, a full disconnect and reconnect is the most reliable way to clear any corrupted state. In Claude.ai, go to Settings, remove the MCP connector entirely, and add it again from scratch using the correct endpoint URL and fresh credentials.
This resets the authorization, the cached tool list, and any session state that may have accumulated. It takes under two minutes for most connectors. With Adsroid MCP, the reconnection process involves pasting the endpoint URL and your API key into the connector settings in Claude.ai. No config file editing, no developer environment required. Once saved, the full tool list loads in the next new conversation.
A full reconnect is not a last resort, it is often the fastest path to a clean working state. If you have been troubleshooting for more than 20 minutes, skip to this step.
Why MCP Tools Sometimes Don’t Show Up Even After a Successful Connection
It is worth separating two different problems that often get conflated. The first is a connection failure, where the AI client cannot reach the MCP server at all. The second is a tool visibility issue, where the connection succeeds but the tool list is empty or incomplete.
Tool visibility problems are usually caused by one of three things: the tool list was cached from a previous session and hasn’t updated, the server returned an error during tool discovery that the client silently ignored, or the user’s account doesn’t have permission to access certain tools.
Starting a new conversation resolves the caching issue. Checking the connector’s authorization scope resolves the permission issue. If neither works, reviewing the server’s tool discovery endpoint directly (if you have access to the server logs) will usually show what went wrong during the initial handshake.
If you are working with AI agents for advertising and want to understand what a fully functioning MCP tool list looks like in practice, the 25 AI agent prompts for auditing Google Ads and Meta Ads shows the kinds of tool calls that should be available when an advertising MCP server is connected correctly.
A Note on MCP Security and Write Actions
One concern that comes up when debugging MCP connections is whether re-authorizing a connector or reconnecting from scratch could trigger unintended actions. With well-designed MCP servers, this should not happen.
Adsroid MCP, for instance, routes all write actions (creating, modifying, or pausing campaigns, ad sets, ads, or budgets) through Claude.ai’s built-in tool confirmation step. The user sees the exact action and its parameters before anything executes. New campaigns and ads are created in a paused state by default. So even if you reconnect mid-troubleshooting, no ad account changes occur without explicit approval.
If you are evaluating whether AI can safely manage a Google Ads account, that article covers the access models, approval workflows, and data handling considerations in more depth.
Frequently Asked Questions
Why is my MCP server not connecting in Claude?
The most common causes are an expired authorization token, an incorrect endpoint URL, or the connector being disabled at the conversation level in Claude.ai. Start by checking the connector status in Settings, re-authorizing if prompted, and confirming the endpoint URL matches the one provided by your MCP server.
Why don’t my MCP tools show up in Claude even though the connector is connected?
Tool lists are fetched at the start of a conversation and cached for its duration. If the server was updated or if the authorization was incomplete, the tool list may be empty or stale. Start a new conversation to force a fresh tool discovery request. If that doesn’t work, disconnect and reconnect the connector entirely.
What does an MCP OAuth error mean?
An MCP OAuth error typically means the authorization token has expired, the OAuth flow was not completed successfully, or the permissions granted during authorization were insufficient. Re-authorize the connector and approve all requested scopes during the OAuth flow.
How do I fix an MCP authentication failed error?
Open the connector settings in your AI client, remove the existing credentials, and re-enter fresh credentials. For API-key-based connectors like Adsroid MCP, this means entering a valid API key in the connector configuration. For OAuth-based connectors, complete the authorization flow again.
Why are only some of my MCP tools showing, not all of them?
Partial tool lists usually indicate a permissions issue or that certain data sources are not connected to your account. Check that you have granted all required scopes during authorization and that any underlying accounts (such as a Google Ads or Meta Ads account) are linked to the platform backing the MCP server.
Can a plan restriction cause MCP tools not to show?
Yes. Claude.ai gates MCP connector access to certain plan tiers. If you are on a free plan or a workspace-managed account with restricted connector permissions, the connector may not load tools. Check your subscription level and, if on a team plan, confirm with your workspace administrator that external MCP connectors are allowed.
Does reconnecting an MCP server cause any unintended changes to my ad accounts?
Not with well-designed MCP servers. Adsroid MCP, for example, requires explicit user confirmation for every write action through Claude.ai’s tool confirmation step. Reconnecting the server resets the connection state but does not trigger any actions in connected ad accounts.
How do I reconnect an MCP server in Claude.ai from scratch?
Go to Settings in Claude.ai, navigate to the Connectors section, remove the existing MCP server entry, then add it again using the correct endpoint URL and your credentials. Start a new conversation after saving to load the fresh tool list.
Why did my MCP tools stop working after a server update?
After a server-side update, the tool list cached in your current conversation becomes outdated. Start a new conversation to trigger a fresh tool discovery request. If new tools still don’t appear, disconnect and reconnect the connector in your AI client settings.
If you are working specifically with advertising accounts and want a connector that covers Google Ads, Meta Ads, Google Analytics 4, and more within a single endpoint, Adsroid MCP is built for exactly that use case and connects to Claude.ai in under two minutes without any developer configuration.
