Academy Use case
Connect Your AI Tools: ChatGPT, Claude, Claude Code and Cursor on One Memory
How AI tools connect to an engine, what each connection can read and save, and how files and your website get in. No key to paste for most.
Once your engine has a purpose, it knows what it is for but holds almost nothing. The next step is to connect the AI tools you already use, so they can start saving what matters and reading it back. The design rule that makes this adoptable is that nobody is asked to feed a new system by hand. You keep working in ChatGPT, Claude, Claude Code or Cursor; they save to the engine as they go, and every one of them reads the same memory.
Context Engine remembers; your AI tools do the thinking. It is not another chat window and it does not run your agents. A connection is simply the door between one AI tool and the engines you choose.
What a connection is
A connection is one signed-in app, or one key. Each has its own address, and each reaches only the engines you tick, at one of two levels:
- Can read: the tool can search the engine and open memories, but not save.
- Can read and save: the tool can also save new memories and edit existing ones.
You can change either at any time under AI tools in the console. The change applies from the tool’s next request, with nothing to reinstall. Everyone else in an engine sees a connection as the app and then whose it is (“Claude Code · Priya”), on Home, in Memories, in a memory’s history and under an answer’s sources. So when something was saved, you always know by which tool and for whom.
Connecting, tool by tool
Start from Connect an AI tool. It is the last step of creating an engine, and it sits at the foot of the sidebar after that. Choose the app, and the screen says in one line what the connection will reach. Change… beside that line is where you pick engines and levels. Then it shows that app’s own steps, with this connection’s own address.
ChatGPT and Claude (on the web or the desktop app) sign in. The screen tells you which menus to open in the app and gives you the address to add. The app sends you to an approval screen, you choose Allow, and you are connected. There is no key to paste anywhere.
Claude Code is one command, shown on the screen with your connection’s address in it:
claude mcp add --scope user --transport http ce-sales https://app.contextengine.com/mcp/c/<id>Then, in Claude Code, type /mcp, choose that server, and choose Authenticate. The server is named after the connection, ce- and its name, so a second connection on the same computer sits beside the first instead of replacing it. --scope user makes it available in every project; Show me how on the screen explains the project-scoped variant, which is how you keep one client’s repository talking only to that client’s engine.
Cursor gets an Add to Cursor link, and the entry for ~/.cursor/mcp.json if the link does not open.
Something else, meaning any agent or script you build yourself, uses a key, because a script has nowhere to click “approve”. Create my key makes a personal key on the spot, reaching the engines you choose at the level you choose. It works over MCP or the HTTP API, sent as a bearer token. The key is shown once, so copy it before you close the dialog. An engine owner can also make an engine key, which belongs to the engine rather than to a person; what it saves is shown as “Engine key”, whoever holds it.
Two status steps on the screen complete by themselves: Signed in turns green when the app first reaches the connection, then First memory when the first save arrives. Send these steps copies everything as text or starts an email, if someone else is setting the tool up. The sign-in itself still has to be yours.
What the tool gets once it is connected
When the AI tool starts a session, it receives a short brief. Here is the shape of one, simplified, for a connection that reaches two engines:
You are connected as "Claude Code". This connection reaches 2
engines, read and write. Every write names one.
ENGINES
- sales: Sales (Read and write)
Our sales relationships: accounts, people, decisions, promises.
Not: internal planning.
- mine: Priya's notes (Read and write)
My own working notes and preferences.
CHANGED SINCE YOUR LAST SESSION
- sales · company::meridian-freight: Meridian Freight
edited by ChatGPT · Sam
- sales · task::security-overview-tomas: Security overview to Tomas
written by Claude · Priya
- mine: nothing changed.
NEVER STORE
Secrets (API keys, tokens, passwords, private keys) ...From there it can search every engine the connection reaches by meaning (each result names its engine), open a memory, and, at Can read and save, save new ones. It is also told what not to keep: not the step-by-step progress of the task in hand, and not a second copy of something already stored. That is what keeps a memory fed by many tools from turning into noise.
Keeping control
Connections are easy to make, so they need to be easy to see and to stop.
- Pause. Any connection can be paused. Every call it makes is refused with the reason until you choose Resume, and it comes back exactly as it was.
- Cut off. Owners and admins see every connection reaching their engine, with its access and last use, and can cut any of them off from that engine.
- See what it saved. From a connection’s card, See what it saved opens Memories on everything it saved in the last seven days. One button undoes it: what it created is archived, and what it edited is put back as it was. Restore undoes the undo.
- History. Every change records who made it, with which tool, and what it was before.
A note on keeping clients apart. Claude on the web and the desktop app switch every connector on in every chat, and ChatGPT switches one on per chat. If you keep clients separate, switch on only that client’s connection in each chat. In Claude Code, use a project-scoped connection per client repository.
Files and your website
Not everything worth remembering comes through a conversation. Two other ways in are worth doing early.
Files. The master services agreement, the pricing sheet, the onboarding checklist, the brand deck: drop files, a whole folder or a .zip on Files in the console. Before anything is read, it shows which engine they go into and every file in the set, so you can untick what should not be there. Text and PDFs become searchable; pictures are described once, so they can be found by what they show. A folder’s name becomes a tag on every file in it. Each file then shows whether it is searchable, and if not, why. Your AI tools find files the same way they find any other memory.
Your website. When you create an engine for an organisation, give it your website. It reads a handful of pages, keeps none of them, and suggests values, voice, a few notes and your colours on Home. Nothing is saved until you keep it.
What there is not: Context Engine does not connect to your email, chat, calendar or CRM directly. If you want an AI tool to work with one of those, use the tool’s own connector for it (Claude and ChatGPT both offer some), and let Context Engine be the memory that tool reads from and saves to. The pre-meeting brief and CRM pieces show that shape.
Tuning it
| Choice | Starting point | Alternatives |
|---|---|---|
| Access | Can read and save for your own tools | Can read for tools that should only look things up |
| Engines per connection | the ones that tool’s work touches | one engine per connection, for strict separation |
| Claude Code scope | --scope user | project-scoped, one per client repository |
| Scripts | a personal key | an engine key, when it should not be anyone’s |
| First files | the ten documents people ask about most | a whole folder or .zip |
How many connections each engine can have depends on your plan; see plans.
Failure modes, and the fixes
The tool connected but saves nothing. Check the level: Can read cannot save. If it is Can read and save, tell the tool once what is worth keeping (“decisions with their reasons, people, commitments”), or put it in its standing instructions.
Memories land in the wrong engine. The tool chooses by each engine’s purpose. Sharpen the purposes, especially what does not belong, and move the misfiled memories; their history goes with them.
A tool saved too much. Open See what it saved and undo it in one step, then tighten the tool’s instructions. Pause it in the meantime if you need to.
Two tools, two versions of the same person. Identical memories are merged with both histories kept, and similar ones are asked about on Home. Answer those, and the person is known once.
What changes
The visible change is that every AI tool you use stops starting from nothing. Ask Claude about a customer you discussed in ChatGPT last week and it knows, and it can tell you which engine the answer came from.
The compounding change is quieter. Every tool you connect makes every other tool better informed, because they all read and save to the same memory. You connect once; the value accrues every session after.
Build this yourself
- Give the engine a purpose first, so your tools know what belongs in it.
- Connect the AI tool you use most, at Can read and save. Ask it to save one thing, and watch First memory turn green.
- Connect the rest of your tools, and pick the engines and level for each.
- Drop your ten most-asked-about documents into Files.
- Open AI tools after a week and use See what it saved to check each tool is saving what you want.
Then read one brain, many agents for why this matters more than the model. To start, sign in free.