AgentCustom connector not working? How to fix it in Claude, ChatGPT, Grok and Perplexity
A custom connector not working almost always traces back to one of five things: the AI company's servers can't reach yours, the sign-in has gone stale, your plan or workspace doesn't allow it, the app is holding an old list of tools, or the menus moved since the guide you're following was written. Here's how to tell which one you've hit, and the fix in Claude, ChatGPT, Grok and Perplexity, taken from each maker's own help pages, read November 13, 2026.
Quick answer
Disconnect the connector and connect it again, or remove it and add it back. That forces a fresh sign-in and a fresh list of tools. If it still fails, check three things: the server address has to be on the public internet, your plan has to allow custom connectors (Claude Free allows one), and on a work account an admin may need to add it first.
Why is my custom connector not working?
Usually because the link between the AI app and the server broke, not the server itself. Match what you see to the likely cause:
| What you see | Likely cause | Try first |
|---|---|---|
| Disconnected, or it times out while you add it | The AI company's servers can't reach the address | Test the address from outside your VPN; use a public address or a tunnel |
| Connected, but no tools show up | The server lists no tools, or the app kept an old list | Refresh the tools, or disconnect and connect again |
| Asks you to sign in again every day | The sign-in isn't renewing | Sign out of extra accounts, reconnect, then tell the server's maker |
| Reads fine, but can't create, send or post | Read-only access from your plan, the sign-in or the server | Reconnect and approve every permission; check your plan |
| The add button is missing | Your plan, a workspace rule, or a menu that moved | Check the plan table below, or ask your admin |
| Connected, but the AI answers without it | It isn't switched on for this chat | Turn it on for the chat, or name it in your question |
The first row is the one people miss. Anthropic's help page says Claude reaches a custom connector from Anthropic's cloud, not from your computer, even in Claude Desktop and Cowork. So a server that opens fine on your laptop can still fail behind a VPN, a company firewall or a home network. Grok's docs and OpenAI's guide likewise ask for an address on the public internet, or a tunnel.
Before you change anything else, work down this list:
- Reload. Hard-refresh the page or restart the desktop app.
- Check the account. Perplexity's troubleshooting page blames many connector problems on being signed in to a different account than the one that owns the connection.
- Reconnect. Disconnect, wait a few seconds, connect again, and finish the sign-in in the window that opens.
- Remove it and add it back. If reconnecting doesn't stick, delete the connector and add it fresh with the same address.
How do you fix a custom connector in Claude?
Start in Customize, then Connectors, and figure out which kind you have: a remote connector added to your Claude account, or a local server written into Claude Desktop's settings file.
Anthropic's help pages, read November 13, 2026, give the basics. On Team and Enterprise, an Owner adds the connector first, then each member clicks Connect. If sign-in fails, disconnect and reconnect from Customize, Connectors. Once it's connected, switch it on for each chat with the + button, then Connectors. To change the address or settings, Anthropic says to remove the connector and add it again.
Claude Desktop says "Server disconnected"
That message usually comes from a local server listed in Claude Desktop's config file, which Anthropic treats as separate from remote connectors. One checklist on X names stray log lines printed where Claude expects messages, relative file paths, and settings missing from the env block. Another post adds that Claude Desktop starts the server on its own, without the shortcuts your terminal knows, so it may not find tools like npx.
Claude Code doesn't list a remote server
Check your project's .mcp.json file. One tip on X: an entry with a url but no "type" is read as a local server and skipped, so add "type": "http". If the first tool call fails after a weekend, the server may have dropped its sign-in, and one fix is the /mcp reconnect command. Another post traced sign-in errors to bugs fixed in newer Claude Code versions, so update first.
Got a connector running? Point the next one at your agents: with AgentMsgr, Claude, ChatGPT, Grok or Perplexity can brief the agents in your group chats and read what they say back. Start free or text the demo room.
Why is my ChatGPT custom connector not working?
Most often it's the setup path, the type of sign-in, or a tool list that didn't refresh.
OpenAI's developer guide, read November 13, 2026, says to use ChatGPT on the web: open Plugins, select the plus button, then Add custom MCP server, and finish with Create as a plugin. Then install it and pick it with @ in a new conversation. Workspace rules, including Lockdown, apply on top.
- The sign-in type. ChatGPT offers OAuth (the "sign in with your account" kind), no sign-in, or a mix. OpenAI's sign-in guide says ChatGPT can't present custom API keys, which matches a builder's warning that a server guarded by a fixed key won't connect.
- An old tool list. When a server changes, open the connection in ChatGPT Plugins, select Refresh, then start a new conversation.
- A 401 error. OpenAI's troubleshooting page ties it to a sign-in that has to start over. Reconnect; if it keeps happening, send the error to the server's maker.
Do you need developer mode for ChatGPT custom connectors?
Posts on X can't agree, and some of the mix-up comes from ChatGPT itself. One user complained that ChatGPT kept saying developer mode was needed because OpenAI's docs lagged; another said you must switch it on at a desktop; a third called one such developer-mode claim "already wrong". OpenAI's developer guide, read November 13, doesn't mention developer mode at all; it sends you to Plugins on the ChatGPT website. The plan list, including which plans get read-only access, is in OpenAI's help center, which wouldn't load for us on November 13. Check it there, not with ChatGPT.
Why is my Grok or Perplexity custom connector not working?
For Grok it's usually reachability or an admin step. For Perplexity it's often a filter, a workspace or a blocked pop-up.
Fixing a custom connector in Grok
The connector docs from SpaceXSI (formerly xAI), last updated October 10, 2026, say you add one at grok.com/connectors: New Connector, then Custom, then the server address and sign-in. On Grok Business and Enterprise, a team admin sets it up in the console before members see it. Grok uses a connector's tools on its own when your question is about that service, so if it seems to ignore one, name the service in your question. The Grok Bot app is a separate product with its own Connect apps screen; its problems are covered in Grok Bot connector not working.
Fixing a custom connector in Perplexity
Perplexity's help center has the most detailed troubleshooting page of the four. The fixes that matter most:
- Connected but ignored. If Sources or Focus is set to Web-only, Academic or Social, Perplexity won't call connected apps. Switch it back, or pick the connector with @.
- Missing in a work space or Project. A connector from your personal account isn't there automatically. Connect it again from inside that space or Project.
- The sign-in window closes. Sign-in needs pop-ups and third-party cookies, which ad blockers often stop. Try a private window with extensions off.
- New tools don't appear. Disconnect and connect again so Perplexity fetches the tool list fresh.
Custom servers do run on Pro: one Perplexity Pro user linked a custom connector to their own agent server and had Perplexity talking to Codex through it.
Why does my connector keep asking me to sign in again?
Because the token that keeps you signed in isn't being renewed, and the cause is often in your browser rather than the server.
One Claude user was asked to sign in to a server about once a day, sometimes mid-session. Perplexity's help center lists the usual causes of loops like that:
- You're signed in to two accounts with the same provider, say two Google accounts, and the token gets renewed against the wrong one.
- You changed a password or a two-step login setting, and the provider cancelled every older token.
- Third-party cookies are blocked, so the app can't renew quietly.
- A work admin removed or restricted the app.
Sign out of the extra accounts, reconnect once, and see if it holds. If it still asks every day with one account signed in, the fault is probably in the server's sign-in setup, and only its maker can fix that.
Which plans allow custom connectors?
All four apps allow them somewhere, but the rules differ, and posts on X often get them wrong.
| App (maker) | Who can add one | Watch for |
|---|---|---|
| Claude (Anthropic) | Free, Pro, Max, Ultra, Team and Enterprise | Free allows one. On Team and Enterprise an Owner adds it first |
| ChatGPT (OpenAI) | Depends on plan and workspace; OpenAI's help center keeps the list | Added on the ChatGPT website; workspace rules and Lockdown apply |
| Grok (SpaceXSI) | All Grok users | On Business and Enterprise an admin sets it up first |
| Perplexity | Some custom servers are limited to Pro, Max or Enterprise | In an organization an admin has to allow members to add them |
Here's how the posts disagree. One setup thread said its connector works on Claude Free (as the one custom connector), Pro and Max, with the owner adding it on Team and Enterprise, which matches Anthropic's page. Another reply, citing a third-party company's docs, listed only Pro, Max, Team and Enterprise, which leaves out Free and Ultra. When two posts clash, the maker's help page with a date on it wins.
Why don't the menus match the guide I'm following?
Because the apps rename menus faster than guides get rewritten. Perplexity is mid-move: its steps point to Plugins in the sidebar, or Customize, then Connectors, for accounts that don't have Plugins yet. Phone apps differ again: one user found the ChatGPT app's path ran through Settings, Plugins and an MCPs button, unlike a vendor's help page. If a label is gone, find the maker's newest help page and follow the screen in front of you.
How do you fix the AgentMsgr custom connector?
Check two things first: the address you added, and the button you picked when you signed in.
AgentMsgr has two addresses. https://app.agentmsgr.com/mcp/user is for your own assistant, so Claude, ChatGPT, Grok or Perplexity can read and post in your chats as you. https://app.agentmsgr.com/mcp, without /user, is for agents that join as members of your chats, and they sign in with a key from a pairing message, not with your email. If an agent is the one that won't join, the guide to pairing an agent with one pasted message has a list of what to check when it stalls.
- It reads but won't post. At sign-in you chose Allow (read and post) or Allow reading only. To switch, tap Disconnect under You, Connected apps, then connect again from your assistant and pick the other button.
- ChatGPT stops using it. Picking AgentMsgr with @ covers one message, so @mention it again when a follow-up needs fresh news from your chats.
- Perplexity skips your chats. Let the sign-in window pop up, and check that the Sources filter isn't set to Web-only.
- Claude Free won't take it. Free allows one custom connector, so AgentMsgr has to be that one.
- Posts pause for a while. Each connection can post 30 times and upload 20 files an hour, on top of your plan's usage.
The setup itself is in the Claude custom connector guide and the guide for ChatGPT, Grok and Perplexity, with every screen on the connect-your-assistant page.
The full walkthrough for this connector, app by app, is in the AgentMsgr connector sign-in guide.
Sources: Anthropic, Get started with custom connectors using remote MCP and Use connectors to extend Claude's capabilities; OpenAI, Add custom MCP server, Connect and test your plugin, Authentication and Troubleshooting; SpaceXSI (formerly xAI), Grok connectors (last updated October 10, 2026); Perplexity, Adding custom remote connectors, Troubleshooting connectors and Local and remote MCPs. All read November 13, 2026; OpenAI's help center page on plans didn't load that day. Posts on X are public posts from October 14 to November 13, 2026, and each was checked against X's public post data on November 13, 2026. The AgentMsgr steps describe the app this month. AgentMsgr is independent and isn't affiliated with Anthropic, OpenAI, SpaceXSI or Perplexity.
Custom connector questions people ask next
Does removing a custom connector delete anything in the service?
No. Removing it ends the link and the sign-in inside the AI app, while your data in the service stays where it was. You sign in again and approve the tools again after adding it back. Anthropic notes that changing a custom connector's details means removing it and adding it back anyway.
Can a custom connector reach a server running on my own computer?
Not as a remote connector. Claude reaches custom connectors from Anthropic's cloud, Grok asks for a tunneling service, and ChatGPT offers its Secure MCP Tunnel. Perplexity runs local servers in its Mac app, and Claude Desktop has local servers of its own in a settings file.
Why does ChatGPT ask me to confirm every time a connector acts?
OpenAI's guide says write actions need your confirmation by default, and ChatGPT treats any tool its server hasn't marked read-only as a write. You can tell it to remember your choice for the rest of that conversation; a new conversation asks again.
Is it safe to add a custom connector someone shared online?
Only if you trust whoever runs the server. Anthropic, OpenAI and Perplexity all say they can't vouch for servers they haven't checked, and Anthropic and OpenAI warn that a bad server can slip in hidden instructions that steer the AI. Approve the narrowest access that does the job.