Some guides in this series are tidy. This one comes with a list of gaps, because Jan's documentation describes remote MCP servers well in one place and says very little about the desktop screen. We write down what is confirmed, then what is not.
What is confirmed
Jan's agent documentation says remote servers that need OAuth are handled automatically. It describes metadata discovery, dynamic registration and an authorisation-code sign-in with PKCE, using a local redirect. That matches how SpendixAI works, the expense tracker we build for India: the client registers itself, you approve access on a consent screen and there is no key to paste.
The console route
- In SpendixAI open Developer → MCP connections and copy the server address.
- In the Jan console run /mcp and press a to add a server.
- Enter a name, choose http as the transport and enter the address when asked for the URL.
- New servers start inactive. Select the entry and press Space to switch it on.
- Select the server, press Enter and choose Authenticate.
- Jan prints a sign-in link and waits for up to five minutes. Open it, sign in to SpendixAI, choose the workspace and a preset.
The value for the URL prompt is the one this guide calls <your SpendixAI server address>. Jan keeps the token in a file named mcp_oauth.json with owner-only permissions. The same configuration file is shared with Jan Desktop. One more rule: if you set an Authorization header yourself, Jan skips OAuth, so leave headers empty.
What we could not confirm
- A desktop screen for adding a remote server. The desktop page we read shows only command-based servers, with fields for Command, Arguments and Environment Variables.
- The names of the form fields beyond name and transport.
- The release in which remote OAuth first appeared. Changelogs mention MCP hardening and inline approval, but not that.
These steps were checked against Jan's own documentation in November 2026. If your version looks different, trust what you see and the current Jan guide.
Approvals and limits
On Jan Desktop, each tool call is approved separately by default. The approval sits inline in the chat and shows the exact arguments. A setting called Allow All MCP Tool Permissions, under Settings and MCP Servers, skips that step, so leave it off for finance records. The agent console has its own tool policy with read-only, deny and allow modes. Read-only is the one to pick.
What this means in practice
If you are comfortable in a terminal, the console route is a complete setup on its own. If you prefer the desktop window, be aware of the gap. Because the configuration file is shared with Jan Desktop, a server added in the console is stored where the desktop app also reads, but we did not confirm how the desktop window shows it, so we are not publishing a menu path for it.
The sign-in link stays valid only for the five-minute wait. If it lapses, run Authenticate again. Jan keeps the connection details out of your way afterwards, and you can end access from SpendixAI whenever you like without changing anything inside Jan.
Confirmed and unconfirmed side by side
| Area | Status | Note |
|---|---|---|
| Console route with /mcp | Confirmed | Add, switch on, then Authenticate |
| Remote OAuth with PKCE | Confirmed | Automatic; no pre-registered credentials |
| Desktop remote screen | Not confirmed | Only command servers are documented |
| Form field names | Partly confirmed | Only name and transport |
| Introducing release | Not confirmed | Not stated |
| Per-call approval | Confirmed | Default on Jan Desktop |
Once connected
Jan's answers come from the model you choose, so verify anything important, such as tax-related totals, in the matching SpendixAI report. Try these first.
- What did I spend on food last month compared with the month before?
- Which merchants did I pay most often this quarter?
- List my recurring charges and what each costs per year.
Useful ideas start with a clearer picture of your own spending.
Explore your reports


