MCP Servers
Store MCP server entries in the Desktop app (not launched yet)
The Desktop app’s MCP panel stores server entries for a future release. The engine does not launch them yet, so the model cannot use them.
Quick Start
- Open the Manage MCP panel
Find the Manage MCP panel in the app.
- Store a server entry
Enter a name, command, and args, then save. This writes the entry to
mcp.json— it does not launch anything. - Enable, disable, or remove a stored entry
The toggle only flips the
enabledflag in storage. Removing deletes the entry. Nothing else changes today — the agent still uses only built-in tools.
How It Works
The app stores servers through a single endpoint (POST /mcp) with an action of add, remove, or toggle, and lists them with GET /mcp. Storage is the only thing that happens — no process is launched.
These fields are recorded so a future release can pick them up:
| Field | Type | Notes |
|---|---|---|
name | string | Required; adding an existing name replaces it |
command | string | The executable a future release will launch |
args | array | Arguments a future release will pass to the command |
enabled | bool | Defaults to true on add; only flips the stored flag today |
Add errors
The Add button refuses bad input and keeps the panel open so you can correct the entry.
- Blank name. The Add button is refused with an in-app dialog reading “A name is required.” — nothing is written to
mcp.jsonand the panel stays open. - Engine rejection. If the engine refuses the add (for example an invalid command), an in-app dialog surfaces the engine’s error message and the panel stays open so you can correct the entry.
The MCP settings overlay dims a toast beneath its scrim, so errors from this panel surface through the in-app dialog instead — that dialog paints above the scrim and stays readable. If neither the dialog nor the write happens on an older build, see Troubleshooting.
Today vs Future
The panel stores data now; a future release wires it to the agent.
| Today | Future release |
|---|---|
| Store an entry (name / command / args / enabled) | Same storage picked up by the engine |
| Entries survive restarts | Enabled entries launched over stdio |
| The agent uses only built-in tools | The agent sees tools from enabled servers |
Best Practices
Use the SDK for MCP that works today
The Desktop panel is storage-only. To give an agent real MCP tools now, use the SDK path in MCP Overview.
Name servers clearly
The name is the key. A clear, unique name prevents an accidental replace when you add a new entry.
Toggle instead of removing
Disable a stored entry you may want back rather than removing it — the entry and its args are preserved for the future release.