Guide

Connecting Knowl to Claude and ChatGPT

Knowl already runs as an MCP server inside Claude Code, Cursor, Codex, Gemini CLI and Claude Desktop — those read a replica on your own machine and need none of this. A connector is for the places where no process of yours can run: claude.ai and chatgpt.com in a browser, and agents in CI containers where syncing a replica to answer one question is waste. Browsers connect with the sign-in flow below; agents that cannot open a browser use a token from your account.

Experimental

This works, and it is running in production against real workspaces. It is also new, and we would rather tell you what that means than let you find out.

What is stable. The server address, the way you connect, and what the four tools do. Memory you record through a connector is stored exactly as memory recorded any other way — nothing here is a separate or throwaway store, and nothing is lost if the feature changes.

What may change. The tool surface: names, arguments, and which tools exist. If it does, a connector may stop working until you remove and re-add it. We do not expect to need that often, but this is not yet something we promise against.

What is already known and will not change soon. The two platforms do not behave the same way. Claude searches your memory on its own; ChatGPT will not until you tell it to, and the fix is one line you add yourself. That is a difference in the platforms rather than in the server, and it is covered under ChatGPT below.

What is not missing any more. Until recently there was no way to see or disconnect what you had authorized. Account → Connected applications now lists every connector on your account with the workspace it reaches and whether it can write, and disconnects any of them — see Revoking.

What a connector can do

It reads one workspace's team memory: the decisions your team made, the findings they verified, and the constraints they work under — recorded by their coding agents as the work happened.

Four tools, and which of them appear depends on the access you grant:

Tool	Access
Search team memory	read
List contradictions	read
Record knowledge	read and write
Correct knowledge	read and write

One workspace per connector, chosen when you connect. No tool takes a workspace argument, so the application cannot reach another one — and if a token ever leaks, exactly one workspace is exposed. To connect a second workspace, add a second connector.

Claude

In claude.ai, open Settings → Connectors → Add custom connector.
For the server URL, enter:
```
https://api.knowl.cloud/mcp
```
Leave Advanced settings empty. Claude registers itself automatically; there is no client ID or secret to paste.
You will be sent to knowl.cloud to sign in, choose a workspace, and choose read or read-and-write access.

That is all. Claude reads the usage instructions the server sends on connection and searches your memory on its own when you ask about your project.

ChatGPT

Custom MCP apps need a paid plan — Plus, Pro, Business, Enterprise or Edu — and are web only. They do not appear in the mobile apps.

Turn on Developer mode: Settings → Security and login → Developer mode, or Settings → Apps → Advanced Settings. The label varies by account. On Business or Enterprise an admin must first enable custom MCP under Workspace Settings → Permissions & Roles → Connected Data.
Open Plugins and press +.
Enter the same server URL: https://api.knowl.cloud/mcp.
Sign in, choose a workspace, choose the access level.

ChatGPT needs one more line, and this is the important part

ChatGPT will not search your memory unless you tell it to. The MCP specification lets a client use the usage instructions a server sends or ignore them, and ChatGPT ignores them. Claude does not — the same server, the same instructions, a different result.

Add this to Settings → Personalization → Custom instructions, or to a Project's instructions if your team works in Projects:

When I ask about my project, my code, or decisions we have made, search Knowl first before answering.

Without it you will need to ask for a search every time. With it, ChatGPT behaves like Claude.

Note also that on Plus and Pro, ChatGPT treats MCP tools more conservatively than on Business and Enterprise — expect searches to work and treat writing as unreliable on the lower tiers. That is their restriction, not ours.

A CI container or a cloud coding agent has no browser to approve a consent screen in. For those, mint a token instead: on knowl.cloud, open Account → Connected applications → Create token. Pick the workspace, name it after the agent that will hold it, choose read or read-and-write, and optionally an expiration.

The token is shown once. Copy it then and store it as a secret where the agent runs — the server keeps only a hash, so nobody can show it to you again. If it is lost, revoke it and create another. Then point the agent at https://api.knowl.cloud/mcp with the token as a bearer credential — most agent platforms have a field for exactly this next to the server URL.

One caution from the field: some agent platforms strip environment variables whose names contain words like KEY or TOKEN before starting an MCP server. If the credential must travel through the environment, prefer a name without those words, or paste it into the platform's own credential field instead.

What to expect once it is connected

Ask about your project in plain language. "What did we decide about billing?" or "Why did we move off that library?" work better than keyword strings — the search is semantic as well as literal.

It will tell you when it found nothing. A question your team has never recorded an answer to returns an explicit statement that nothing matched, rather than the three closest entries dressed up as findings. If you see that, the memory genuinely holds nothing on the subject.

Writing is deliberate. If you granted read-and-write, anything the assistant records is marked as having come from a connector. It writes without having run your tests or read your code, so what it stores is worth reviewing — that is why read is the default and write is a choice you have to make on the consent screen.

Revoking

On knowl.cloud, open Account → Connected applications. Each row shows the application, the workspace it reaches, whether it can write, and when it last called. Press Revoke.

Removing the connector inside Claude or ChatGPT is not the same thing, and the difference matters if you think something has gone wrong. Removing it there stops that application choosing to call us, which is fine for tidying up. The credential it was given stays valid. Revoking here destroys the credential.

Once revoked, the next call fails immediately; there is no grace period.

One name on that screen is not ours. An application registers itself when you connect it, and the name it registers is its own claim — anything not verified by us is labelled as such. What is ours, and what is worth reading before you decide, is the workspace and the access level beside it. Nothing can reach a workspace that is not named on its own row.

Revoking does not delete anything the connector recorded. If it stored something wrong, correct it the way you would correct any atom.

When something goes wrong

"This redirect is not registered" — the platform sent a callback URL we do not accept. Registration is limited to claude.ai, claude.com, chatgpt.com and chat.openai.com over HTTPS. If you are seeing this on one of those, tell us the URL in the error; it names the value it rejected.

Authorization failed with a support reference — the OAuth handshake failed somewhere inside the platform. The error will not say where, because the flow runs on their infrastructure. Send us the reference and roughly when it happened.

The assistant does not use Knowl — on ChatGPT, add the custom instruction above. On Claude, disconnect and reconnect the connector so it re-reads the server's instructions.