Connect Claude Code
Point Claude Code at the companion's MCP server and give it the device — the endpoint, the token, the approval, the skill.
Claude Code reaches your device the same way any agent does: through the companion. You copy one connect line, approve the client once, and install a skill that teaches the agent how to drive the device. After that it can show screens, read a button press back, and put apps on the device.
Everything you need is on one page in the companion: Settings → MCP.
Copy the connect line
On the MCP page, open the Connect a client card and pick the Claude Code tab. Press Copy. That mints a fresh access token and drops it into a ready-to-run command:
claude mcp add zeph --transport http <your-url> --header "Authorization: Bearer <your-token>"Run it in a terminal. Add --scope user if you want Zeph available in every project, not just this one.
The token appears once.
Copying is what creates it — it lives only in the snippet you just copied. If you lose it, copy again for a new one; the old one keeps working until you revoke it under Connected clients.
Choose what it may request
Under Access this client may request are five grants, all on by default:
| Grant | What it lets the agent do |
|---|---|
| Device control | See devices, change settings, show screens, hear button presses. |
| Install apps | Add and configure apps and glances on the device. |
| Urgent alerts | Post full-screen alerts that wake the device. Off, and they arrive as quiet notices instead. |
| Read library | See your installed providers, tools, apps and skills. |
| Manage library | Install, configure and remove library items on your behalf. |
Uncheck anything you would rather not hand over before you copy. You can widen or narrow it later from the client's row.
Approve it the first time it connects
The token alone is not trust. The first time Claude Code calls in, the companion asks "Let Claude Code use your Zeph?" and lists exactly the grants above. Press Allow. Until you do, every call is refused — a token you created is not yet a token you approved.
Install the base skill
Claude Code can now reach the device, but it does not yet know the device's shape. On the same page, under Base skill, press Install to your tools… and pick Claude Code. That copies the zeph-device skill into the agent's skills folder — it teaches it to push screens, ask questions, and read the button you press back.
What it can do now
From here the agent drives the device over that one approved connection. It can manage the device end to end, put a screen up and wait for your answer — the device draws a question with one option per button, and the button you press comes back to the agent that asked. It can also install a published app that then runs on its own, with your Mac closed.
If Claude Code goes away, the last screen it drew goes stale.
Live screens are computed on your Mac. Anything that must keep working while your Mac sleeps belongs in an installed app, in the device's own flash.
The endpoint, if you are wiring it up by hand
The companion's MCP server listens only on your own machine, over local HTTP, and every call carries the bearer token you minted. Any MCP client that speaks streamable HTTP works the same way — the Cursor, Claude Desktop and Other tabs on the MCP page give you the same URL and token in each client's own config shape.