GBrain gives an assistant a door to the things you have connected: your mail, your calendar, web search. You decide which assistant gets a door, what it may reach through it, and how much it may do in a day. Nothing reaches anything until you say so, and everything that happens is on a record you can read. ## The three words **Application** is something you connected: a Google account, an Exa key, the memory of one of your workspaces. It belongs to you, not to any one assistant. **Client** is the access you granted: which applications, at what level, for how long. A client is what you turn up or down. **Connection** is one thing plugged into a client: ChatGPT on the web, Claude on your desktop, a terminal. Several can share one client, and each has its own token and its own switch. So: applications are what exists, a client is what is allowed, and a connection is what is plugged in.

One client sits between what you own and what you plug in. Change a permission once and every connection on that client follows, from its next call.

Why the middle box exists: you will plug the same access into more than one place, and you want one switch for all of them. Turning ChatGPT's access down should not mean hunting through three apps. ## Connecting one Open **Clients**, pick the client you want, then **Add a connection**. ChatGPT, Claude and Claude Code sign themselves in: you paste one address, they send you back here to approve it, and it appears in your clients. Everything else takes a token you copy once.

The Add a connection page. Apps that sign themselves in: ChatGPT, Claude, Claude Code. Everything else, with a token: Codex, Cursor, a workspace of yours, a terminal. — **Pick what you are plugging in.** The next page has the exact lines to paste for that app, including where its settings hide the MCP option.

The address is the same for every assistant: ``` https://gbrain.io/mcp ``` Each app's page has the exact steps for that app, including where its settings hide the MCP option. ## Your workspace's memory Every workspace keeps a memory: the facts, pages and links it has gathered. You can add that memory to a client like any other application, from the client's Applications page or from **Applications**. Set it to **Read** and the assistant can search and recall; set it to **Full** and it can also remember new things, link pages, and forget. The memory arrives as the tools gbrain itself defines: `recall`, `remember`, `search`, `fetch`, and the rest. A client holds one workspace's memory, so a second workspace means a second client. The assistant is told, when it connects, which workspace this is and when to reach for it: at the start of a session and after its context is compacted, to warm up; before answering questions about what is known; and to save something durable you say about yourself or your work. Nothing runs inside the assistant for this to work. It is all the same door. Two things to know: - Nothing listens on the workspace until you add its memory to a client. Remove it from every client and the workspace's memory server switches off again. - Every memory call is on the client's Activity page, with the outcome, the same as every other call. If you are starting from the workspace side, [Use your memory from other assistants](/docs/workspace/memory-anywhere) walks the same ground in four steps. ### Per assistant - **Claude Code.** `claude mcp add gbrain -t http https://gbrain.io/mcp` and sign in when it asks, or pass a token with `--header "Authorization: Bearer "`. The workspace's tools appear alongside the rest, under the server name you gave. - **Claude on the web or desktop.** Add `https://gbrain.io/mcp` as a connector under Settings and approve the sign-in. Same tools. - **ChatGPT.** Add the same address as a connector and approve the sign-in, or install the GBrain app from the directory when it is listed. Deep research works too: the workspace's `search` and `fetch` are the tools it looks for. - **OpenClaw.** Add the address under `mcp.servers` with `transport: "streamable-http"` (its default is the older SSE) and `auth: "oauth"`. For a shared workspace that several people talk to, set `oauth.identity: "per-requester"` so each person signs in as themselves. - **Hermes.** Add the address under `mcp_servers`; it signs in with OAuth on its own. Its `identity_header` can name which profile is calling. ## What the assistant can work out on its own Once connected, it can answer questions about its own access without asking you: - what it is plugged into, and when that access stops working - which tools it can use right now, and which need permission it does not have - what it has already done, and what that cost - how much of today's allowance is left on each tool When it runs into something it cannot do, it will say which permission it needs and give you the page to change it. It cannot change its own access. Only you can, on that page. ## When something stops working An assistant that has stopped reaching your things is usually one of three states, and it will tell you which: - **Switched off.** Someone turned that connection off. Turn it back on from the client's page. - **Term ran out.** A client expires so nothing stays connected forever. Give it more time from the same page. - **At a daily limit.** One tool used up its allowance for the day. The assistant will hand you a link to that limit, where you can add room for today or change the standing number. Everything else it can do keeps working while one tool is paused. ## The rest of this section - **[Applications](/docs/tools/applications)**: the accounts and keys your clients reach through. - **[Permissions](/docs/tools/permissions)**: what a client may do, per application, at a level you set. - **[Limits](/docs/tools/limits)**: the daily allowance that stops a job that runs away. - **[Activity](/docs/tools/activity)**: every call, what it cost, and which connection made it.