Connecting an agent
MemGym speaks the Model Context Protocol (MCP), so an AI agent — Claude Code, or Claude with a custom connector — can work with your deck directly: read your cards, write new ones from something you are reading, and record real practice reps.
This is available in account mode (signed in at memgym.com). The connection is to your deck, and the agent acts as you.
What it is, in plain language
Your agent gets a scoped key to your deck. With it, it can see your cards, add cards, record practice outcomes, and organise the queue — the same things you can do in the app, through the same rules. It cannot read your email, change your password, or sign in as you anywhere else.
The most useful thing about it is not automation. It is that the thing you are already talking to about a topic can turn that conversation into cards, and practise them with you later.
Setting it up
Claude Code
claude mcp add --transport http memgym https://memgym.com/api/mcp
Then use it. The first time the agent calls MemGym, your browser opens a MemGym consent screen.
Claude (claude.ai)
Add MemGym as a custom connector using the same URL — https://memgym.com/api/mcp — over HTTP transport. The consent step is the same.
The consent screen
You will land on memgym.com/authorize, already signed in, and be asked to approve four things:
- View your deck and collections
- Add cards
- Record practice outcomes
- Organise and retire cards
Approve, and you are connected. There is no token to generate, copy, or paste anywhere — signing in to MemGym in your browser is the authentication step. If a set of instructions anywhere tells you to hunt for an API key, you are reading about a different product.
What the agent can actually do
Seven tools, and this is the whole list:
| Tool | In plain terms |
|---|---|
list_deck |
See your cards — optionally just the ones that are due, or in the backlog |
get_card |
Read one card in full, including its prediction history |
list_collections |
See your collections and folders |
add_card |
Write a new card into your due feed |
record_outcome |
Record one completed predict-then-test rep |
promote_card |
Move a backlog card into the due feed |
retire_card |
Take a card out of rotation |
There is no tool that deletes a card, edits your account, or clears your history. An agent cannot reach those.
The contract still holds
This is the part worth understanding before you let an agent practise with you.
record_outcome takes both a predicted score and an outcome score, and both are required. predicted is how well you expected to do before checking the answer; outcome is what actually happened after. Each is 1 (missed), 3 (partial) or 5 (nailed) — the same three values as the buttons in the app.
So an agent cannot record "I got it right." It can only record a prediction and a result, and the gap between them is what MemGym measures. There is no way through the agent path to log a rep without a prediction, which means the numbers on your calibration page mean the same thing whether you practised in the app or through an agent.
A good prompt for practising this way is therefore: ask me the card, take my answer, ask me how confident I was before I saw the answer, then record both. An agent that predicts on your behalf is guessing at the one number that only you can supply.
Access and revocation, honestly
- The token your agent receives is scoped to your deck and expires after 7 days, after which the agent reconnects through the same consent screen.
- There is no revoke button yet. If you want to cut an agent off sooner than that, this version cannot do it — the practical options are to remove the connection at the agent's end so it stops calling, and to know that the token dies on its own within the week. A proper revocation list is not built.
- The authorization step itself is short-lived: the code exchanged during connection is valid for 60 seconds.
That is the current, real state. It is written here rather than left out because a security property you have to infer is not one you can rely on.
What this is not, yet
MemGym has one account per deck today. Connecting an agent connects it to your deck, as you — there is no separate agent identity, no per-agent permissions, and no shared or delegated access. If you are thinking about giving somebody else's agent access to your cards, that is not something this version can express.