The Llamatik app can use a Llamatik Server for generation instead of (or alongside) a local model. This is useful when the device is too slow or the model too large, or when you want to share one powerful machine between several devices. The other side can be a standalone Llamatik Server or another device running the app with Share this device turned on.

Heads-up: connections are not encrypted yet. Messages sent to the server travel over plain HTTP, so they are not 100% secure and anyone on the same network could potentially read them. Be careful about what you send, and avoid sharing passwords or sensitive personal information. Encrypted (HTTPS) connections are planned; both encrypted and unencrypted modes will be supported.

Connect from the app

Open Settings → Connect to Server:

• Discover on network — scan the local network for nodes advertising the _llamatik._tcp mDNS service and tap Connect. Llamatik Server advertises itself automatically (see Configuration); if it doesn’t show up, add it manually. mDNS needs the phone and server on the same network, and it usually doesn’t work across Docker’s default bridge network, VPNs or guest Wi-Fi.
• Add manually — enter the IP address, port and an optional name, then tap Connect.
• Saved servers — servers you connected to before can be reconnected with one tap.

When you connect to a device that is sharing itself, the host owner has to approve you first (see Share this device). Before the first connection to any server the app shows an unencrypted connection warning that you must confirm. The same warning is shown as a card at the top of the server screen.

Once connected, the server card lists its available capabilities (text, image, vision, video, audio), based on the models currently loaded on the server. The Settings screen also shows Connected · <name> (<host:port>) under the Connect to Server row.

Chat without a local model

If you skip downloading a model and connect a server, the chat screen shows the normal input box and sends prompts to the server. The No Model Selected view only appears when there is no local model and no connected server.

Responses are streamed (Server-Sent Events), so text appears token by token.

Status indicator

While a server is connected, the chat toolbar shows a status dot. The app checks the server every 10 seconds:

The same check also refreshes the capability list, so models loaded on the server after you connected show up without reconnecting.

Fallback to local generation

If the server is unreachable, has no text model loaded, times out or fails mid-request, the app falls back to local generation when a local model is available. When the request already failed, a dialog offers to retry the same prompt locally.

Using the client from code

The :llamatik-network module provides LlamatikServerClient, the same client the app uses:

val client = LlamatikServerClient(LlamatikServerConfig(host = "192.168.1.20", port = 8080))

if (client.isReachable()) {
    val status = client.getModelsStatus()           // what is loaded on the server
    client.streamGenerateWithContext(
        systemPrompt = "You are a helpful assistant.",
        contextBlock = "",
        userPrompt = "Explain Kotlin Multiplatform in one sentence.",
    ).collect { chunk -> print(chunk) }              // streamed token by token
}

Notes:

• The default HTTP client sends Content-Type: application/json. If you pass your own HttpClient, install ContentNegotiation with JSON and set that default content type, otherwise requests fail with Fail to prepare request body for sending.
• Streaming calls must read the response as it arrives (preparePost(...).execute { }). A plain post(...) buffers the whole body first and defeats streaming. The built-in stream* functions already do this.

Web (Wasm) clients

Browsers enforce CORS. Llamatik Server allows the Content-Type header for cross-origin requests; make sure you run a server version that includes it, otherwise the preflight (OPTIONS) request is rejected with 403 Forbidden and the app reports that it cannot connect.

Also note that a page served over HTTPS cannot call a plain http:// server (mixed content). Until encrypted connections are available, use a server on the same scheme as the page, or a TLS reverse proxy — see Deployment.