Vibium Station lets people and AI agents use real devices from anywhere: see them, touch them, and talk to them.
Each device sits on a rig: today, a Raspberry Pi running Roadie with a phone plugged in. The rig dials out to the station, so it works from behind any router.
flowchart TB
P["Web browser"]
A["Agents<br/>with an API token"]
C["Chat<br/>an LLM on the station<br/>with Vibium (MCP)<br/>and voice"]
S["Station<br/>the hub"]
R1["Rig<br/>Roadie on a Pi"]
R2["Rig<br/>Roadie on a Pi"]
R3["Rig<br/>Roadie on a Mac"]
D1["iPhone"]
D2["Android phone"]
D3["iPad"]
P --> S
A --> S
C --> S
S <-- "tunnel" --> R1
S <-- "tunnel" --> R2
S <-- "tunnel" --> R3
R1 --- D1
R2 --- D2
R3 --- D3
- People open a device in the browser and drive it.
- Agents drive it through the API, with REST and WebDriver BiDi.
- Chat runs on the station: you type what you want done, and an LLM you set up does it on the device through Vibium's MCP tools, speaking and listening with text-to-speech and speech-to-text models. The station's pages call each chat a jam session.
- Video, sound and input go over WebRTC, directly or through the station's TURN relay.
- Each rig dials out to the station, which is the hub, and checks a grant signed by the station before it takes any input.
A station carries each device's sound in both directions at once (full duplex), so people and agents can talk to a device and hear it answer. That makes it a place to test voice assistants, voice agents and any app where sound matters.
- In the viewer, you hear the device and talk to it with your own microphone.
- Agents speak and listen with Vibium:
vibium sayspeaks into the device, andvibium listenturns its answer into text, with how long it took (say and listen). - A chat can talk to the device and check what it says back. Speech can run on the station itself, with Piper speaking and sherpa-onnx listening, or in the cloud.
- Recordings keep both sides of the sound, and the Record Store plays them.
- Sign-in with a password and an authenticator app, or a tailnet identity
- Users and teams: regular users see only the devices their teams were given
- Accounts that expire on a date, for a trial or a demo
- One person or agent in control of a device at a time, and one watching
- Chat: an LLM you set up does what you ask on a device
- Two-way sound, with speech on the station (Piper and sherpa-onnx) or in the cloud
- A Record Store for recordings, screenshots and sound, played in place
- Runs on your LAN, your tailnet, or the public internet
No release is published yet, so build it with Go 1.27.4 or newer:
go build ./cmd/station
./station passwd # the owner's password
./station # LAN: https://<your ip>:8443
./station --tailnet --allow you@example.com # or your tailnet
./station --public station.example.com # or the internetThen:
- Make an enrollment code on Settings → Rigs.
- On the rig, run
roadie enroll https://<station> --code <code>. - Open the phone from Devices.
- For chat, add an AI model on Settings → AI.
| To | Read |
|---|---|
| Run a station on your own server | Self-host a station |
| Run one on Fly.io | Deploy to Fly.io |
| Sign in, enroll a rig and check it all works | First steps |
| Set up chat and voices | Set Up Jams |
| Point an agent at a device | Reach a device through the station |
| Use the API | API.md |
| Look up a command or flag | Commands and flags |
| See how it keeps devices safe | Security in short, Nostr keys |
| Embed it in a service of your own | Packages |
| Read the design | Remote spec, teams, jams |
Milestones R0 to R5 of the remote spec are done. Packaging and hardening (R6) and recordings (R7) are in review.