Claude connectors missing from settings, stuck on Connect, or throwing errors mid chat? This article sorts the causes by symptom: plan limits, admin approval, sign-in loops, unreachable servers, bad config files, and tool errors, each with a concrete fix.
Few things in Claude feel as annoying as a connector that worked yesterday and has vanished today. The Connectors menu is empty, the Connect button spins forever, or a tool that looked healthy throws an error halfway through a chat. The good news is that nearly every case traces back to one of a handful of causes, and you can rule them out in about ten minutes if you go in the right order.
This article follows that order. It starts with the boring checks (plan, admin settings, account) because they explain most "not showing" reports, then moves on to sign-in failures, server reachability, local config mistakes, and errors that only appear after a connector shows as connected. Menus and labels move around often, so if a button is not where you expect it, look for the Connectors or Customize area in settings.
💡 Rule of thumb: If the connector is not visible, suspect your plan or an admin setting. If it is visible but will not connect, suspect sign-in or reachability. If it connects but fails in chat, suspect the tool call itself.
Start by naming your symptom, because each one points to a different fix:
Connectors, and custom connectors that point at a remote MCP server in particular, belong to paid plans. On a plan without them, the Connectors option simply does not appear, and no refresh will bring it back. Open your account settings, confirm which plan is active, and compare it against the official connectors help article. If you recently upgraded or switched workspaces, sign out and back in so the new plan is picked up.
On Team and Enterprise plans, an Owner or Primary Owner has to enable a connector for the organization before anyone can use it. In the admin area that means opening the connector list through Browse connectors and choosing Add to your team. Enabling a connector only makes it available. Each person still has to sign in to it individually, so a colleague seeing "connected" tells you nothing about your own account.
Three questions worth putting to your admin:
These take two minutes and fix a surprising share of cases. Do them before you touch any server settings.
Stale sessions cause ghost problems. Work through these in order and test after each step:
A connector can be connected and still be switched off for the conversation you are in. Open the tools menu near the message box and make sure the connector is toggled on. Then check its tool permissions: individual tools can be set to ask first, always allow, or stay blocked. A blocked tool looks exactly like a missing one.
| Symptom | Most likely cause | Fastest fix |
|---|---|---|
| No Connectors option in settings | Your plan does not include it | Check the plan, upgrade, or ask your admin |
| Connector exists for others, not for you | Never signed in, or not added by an Owner | Ask the Owner to add it, then sign in yourself |
| Connect button does nothing | Blocked pop-up or stale session | Allow pop-ups, hard refresh, retry |
| Connected, but Claude ignores it | Switched off in this chat | Enable it in the tools menu |
| Worked yesterday, fails today | Expired or revoked token | Disconnect, then connect again |
When the connector is visible but sign-in fails, the problem sits somewhere between your browser, the other service's login page, and Claude.
Connecting opens a login page from the other service: your calendar, your CRM, or your own server. If that window never appears, closes instantly, or loops back to the start:
💡 Reconnect cleanly. Remove the connector, close the tab, sign in again, and add it fresh. A half-finished earlier attempt can leave a broken state that repeated retries never repair.
"Authorization with the MCP server failed" usually appears after sign-in has already started. Anthropic's connector troubleshooting documentation lists these server-side causes:
If you only use someone else's connector, send them this list. If you run the server, the next section is yours.
This message sounds like a network outage, yet the cause is often quieter. The important fact: remote connectors on claude.ai run on Anthropic's infrastructure, not on your computer. Claude resolves your server's hostname from the public internet, and if the result is not globally routable, it rejects the connection before a single request is sent. Your access logs show nothing at all.
Claude rejects hostnames that resolve to private ranges (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16), carrier-grade NAT addresses (100.64.0.0/10), loopback or link-local addresses, or a mix of public and private results. Connectors are also IPv4 only, so a hostname with nothing but AAAA records fails.
That explains the classic complaint: "It works in Claude Code and curl, but not on claude.ai." Those tools connect from your machine, while claude.ai connects from the cloud. If your hostname resolves one way inside your network and another outside (split-horizon DNS), or hides behind a VPN, a dynamic DNS provider, or a home router with CGNAT, the web app never receives a usable address.
Check it from a network outside your own:
dig +short your-server.example.com
Every address returned must be public. For a server on your laptop, expose it through a public tunnel or reverse proxy instead of a localhost URL.
With correct DNS, a CDN, WAF, or rate limiter can still reject the request before your application sees it. Look for 403 or 429 responses in your edge logs during a Connect attempt, then allowlist the outbound range published on Anthropic's IP address page, or exempt the MCP and sign-in paths from the blocking rule.
Then check for redirects:
curl -sI https://your-server.example.com/mcp
If the response is a 301, 302, 307, or 308 pointing at a different host, register that target URL instead. Local clients fail on the redirect right away, while claude.ai follows it, drops the credential, and reports an authorization error later.
When a server needs sign-in, Claude first looks up its OAuth metadata. Request each document from a public network and expect a 200 with valid JSON:
curl -i https://your-server.example.com/.well-known/oauth-protected-resource
curl -i https://your-server.example.com/.well-known/oauth-authorization-server
curl -i https://your-server.example.com/.well-known/openid-configuration
Only one of the last two has to answer. The metadata must also offer a way to register Claude as a client (dynamic client registration, client ID metadata documents, or pre-registered credentials) and advertise S256 PKCE support. A proxy that drops the WWW-Authenticate header, or returns 403 on /.well-known/ paths, produces the same "couldn't reach" message with a perfectly healthy server behind it.
Microsoft Entra deserves its own note. If the token request fails with AADSTS9010010, Entra is rejecting the resource value Claude sends. Register your full MCP server URL as an Application ID URI on the API app registration, using a verified custom domain if you are on a platform hostname.
💡 Save the reference ID. When a connection fails on claude.ai, the error page URL contains a value that starts with
ofid_. Copy it right away, because it expires, and include it when you file an issue on the anthropics/claude-ai-mcp tracker.
Servers that run on your own machine take a different path. The desktop app launches them from a config file, so nothing travels over the internet and none of the DNS advice above applies.
The most common local failures are small:
npx or node works in a terminal but not from the app. Use the absolute path to the executable.Run the file through a JSON validator before you restart.
The desktop app writes MCP logs to %APPDATA%\Claude\logs on Windows and ~/Library/Logs/Claude on macOS, with a general MCP log plus one file per server. A server that crashes on startup usually prints the reason there: a missing package, the wrong Node version, or an absent environment variable. In Claude Code, claude mcp list shows the status of every configured server, and /mcp inside a session shows details and lets you sign in.
"Unexpected error while invoking tool", followed by a tool name, means the connection itself is healthy. Claude's request reached the server and the server answered with an error. There is no ofid_ reference ID here, because nothing failed at connection time.
Narrow it down in this order:
Rate limits are another frequent culprit, and they look random. Many services cap how much work runs at once. PicassoIA's own MCP connection is a good example: its image, image editing and video models share a limit of 5 predictions running at once per account, across every credential and MCP connection. At the time of writing, you add that connection from the MCP page of your account, which needs a login. If Claude fires ten generations in one request, some will fail. Ask it to work in smaller batches.
Here is a quick decoder for the messages you will meet most often:
| Message | What it means | Where to look |
|---|---|---|
| Couldn't reach the MCP server | The handshake failed | DNS, firewall, redirects, OAuth metadata |
| Authorization with the MCP server failed | Sign-in started but did not finish | Issuer, audience, PKCE, token speed, redirects |
| Unexpected error while invoking tool | Connected, but the tool call failed | Server logs, MCP Inspector, rate limits |
| No Connectors option at all | Plan or organization setting | Plan page, admin settings |
A broken connector should not stop your day. While you fix it, paste the error text, a log excerpt, or a screenshot into a model that does not depend on any connector. Claude Sonnet 5 is available on PicassoIA, handles code and tool-heavy tasks, and accepts image input, which suits screenshots of error pages.
💡 Strip secrets first. Remove tokens, client secrets, and passwords from any config or log before you paste it anywhere.
Before you close the settings tab, run this final checklist:
ofid_ reference ID is saved if you need support
Once your connectors behave again, give them something creative to do. Picasso IA turns a short prompt into a photorealistic image in seconds with PicassoIA Image, lets you refine the result with PicassoIA Image Editor Pro, and animates a still with PicassoIA Video. Open the platform, write one prompt about a scene you have in mind, change the lighting or the camera angle, and run it again. Ten minutes of experimenting will show you what each model does best, and you can browse every option on the all models page.
