Turn a Strix Halo or Ryzen AI Max box into a model endpoint

Your Strix Halo box already runs the model. The hard part is reaching it from your laptop, editor, or coding agent without opening inbound ports. Run the Tokios connector next to your runtime and you get one authenticated endpoint at https://api.tokios.com that speaks both OpenAI and Anthropic request formats, so Claude Code, opencode, or any OpenAI-compatible client can reach the box without touching firewall rules.

After it runs: the endpoint gap

Framework Desktop, mini-PC, and other AMD Ryzen AI Max+ 395 (“Strix Halo”) owners buy the 128GB unified-memory configuration specifically to run large local models. On this silicon, the community’s general experience is that llama.cpp with the Vulkan backend is more reliable than ROCm/HIP. Once the model answers on localhost, the next thing most people want is to plug in a coding agent — opencode, Claude Code, or a similar tool — against it. That means:

An OpenAI-compatible endpoint reachable from elsewhere. The runtime binds to 127.0.0.1 on the Strix Halo box by default. Reaching it from a laptop or a coding agent running elsewhere means binding to 0.0.0.0, opening a port, or standing up a tunnel — each with its own setup and no per-client auth in front of it.
API shape. Claude Code speaks the Anthropic Messages format; llama.cpp, Ollama, and LM Studio speak OpenAI Chat Completions. Bridging the two usually means writing and maintaining a small proxy.
A stable place to point tools. Swapping between llama.cpp, Ollama, or LM Studio on the same box, or changing which model is loaded, shouldn’t mean reconfiguring every client that talks to it.

Tokios is the connect layer that sits on top of the runtime you already run — it does not replace llama.cpp, Ollama, LM Studio, vLLM, or Lemonade Server.

2 Strix Halo boxes can have enough aggregate memory for DeepSeek-V4-Flash-0731 GGUF weights, but the required llama.cpp RPC path is fragile, insecure, and proof-of-concept. Do not treat it as a production deployment. See DeepSeek V4 Flash local hardware for the experimental boundary.

Set it up on your Strix Halo box

Start your runtime on the box

Run your model server as you normally would, for example with llama.cpp on the Vulkan backend:

llama-server -m your-model.gguf --port 8080

or with LM Studio’s local server, or Ollama:

ollama run llama3.1

Lemonade Server is another option — it auto-detects your Strix Halo hardware and selects the Vulkan backend for you, so you skip the manual backend selection:

lemonade pull your-model
lemonade run your-model

Lemonade Server listens on port 13305 by default and exposes an OpenAI-compatible API at /v1. It also supports AMD Ryzen AI NPUs, which no other backend here does. See Lemonade Server remote access for the full setup guide.Confirm it answers on its own OpenAI-compatible /v1 API before adding the connector.

Get a claim code and pair the connector

In the dashboard, open the Setup tab and select Start pairing to get a one-time claim code shaped TKS-XXXX-XXXX (valid for about 15 minutes).On Linux:

curl -fsSL https://tokios.com/downloads/install/linux | sh -s -- --code TKS-XXXX-XXXX

On Windows (PowerShell):

irm https://tokios.com/downloads/install/windows | iex; tokios-install -Code 'TKS-XXXX-XXXX'

Back in the dashboard, approve the device on the Setup or Connectors tab. The connector’s token is delivered over the tunnel — you never paste it into a file.

Point the connector at your runtime's /v1 endpoint

The connector’s upstream is the runtime’s own OpenAI-compatible API on the box, not a port you expose externally. Default ports: llama.cpp 8080, LM Studio 1234, Ollama 11434, Lemonade Server 13305 (all with /v1 appended; for Ollama use its /v1 OpenAI-compatible API, not the native /api). No inbound firewall rule is needed — the connector only dials out.

On the Models tab, map your upstream model id to a public deployment name such as strix-coder. Clients send this deployment name in the model field, not the upstream id.

Create an API key

On the Keys tab, create a key. Keys are sk-tok-… Bearer tokens, shown once, and can be scoped to specific deployment name patterns.

Use it from your coding agent

Once the connector is paired and a deployment is registered, point your coding agent at Tokios instead of the box’s local address.

export ANTHROPIC_BASE_URL=https://api.tokios.com
export ANTHROPIC_API_KEY=sk-tok-YOUR_KEY
claude --model strix-coder

from openai import OpenAI

client = OpenAI(
    base_url="https://api.tokios.com/v1",
    api_key="sk-tok-YOUR_KEY",
)

response = client.chat.completions.create(
    model="strix-coder",
    messages=[{"role": "user", "content": "Hello from my laptop"}],
)

Claude Code’s model picker only lists claude* ids, so your Tokios deployment won’t appear in the list — pass --model strix-coder explicitly.

If you run more than one backend on the box — say llama.cpp for one model and Lemonade Server or Ollama for another — use the connector config file to define multiple routes and upstreams instead of the single-upstream pairing flow. See Model routing for how to register and switch between several deployments behind the same endpoint.

Why one endpoint

Swap backends without touching agent config. Move from llama.cpp to LM Studio or Ollama on the same box, and every agent that already targets strix-coder keeps working — you only change the connector’s upstream, not the client.
Scoped keys instead of a shared bind address. Give a teammate a key limited to specific deployment patterns instead of opening the runtime’s port to your network.
No inbound ports on the box. The connector only makes outbound connections, so you don’t bind the runtime to 0.0.0.0 or maintain your own tunnel just to reach it from a laptop or coding agent.

Can Strix Halo run Kimi K3?

No. Strix Halo’s 128 GB unified memory pool (~90–110 GB available after OS overhead) is far below what Kimi K3 requires. K3 has 2.8 trillion total parameters and needs about 1.4 TB at its native MXFP4 precision (4 bits per parameter) — before block scales, runtime buffers, activations, and KV cache. Even if you could use every byte of the 128 GB pool for weights, you would still be roughly 1.27 TB short. Moonshot recommends serving K3 on a supernode with 64 or more accelerators. For the full memory calculation and cluster sizing, see the Kimi K3 hardware deep dive.

Looking for a model that actually fits Strix Halo? See what to run instead of Kimi K3 — the best coding and general model for 128 GB unified memory.

Next steps

Quickstart

Go from zero to a working connector, deployment, and API key.

Install the connector

Full connector install and pairing reference for all platforms.

Use Claude Code

Configure ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY in detail.

Connector config reference

Serve multiple models from one box with tokios-connector.json.

Hardware & fit

Serve & connect your runtime

Connect a client or app

Compare

Turn a Strix Halo or Ryzen AI Max box into a model endpoint

After it runs: the endpoint gap

Set it up on your Strix Halo box

Use it from your coding agent

Why one endpoint

Can Strix Halo run Kimi K3?

Next steps

Quickstart

Install the connector

Use Claude Code

Connector config reference

​After it runs: the endpoint gap

​Set it up on your Strix Halo box

​Use it from your coding agent

​Why one endpoint

​Can Strix Halo run Kimi K3?

​Next steps

Quickstart

Install the connector

Use Claude Code

Connector config reference

After it runs: the endpoint gap

Set it up on your Strix Halo box

Use it from your coding agent

Why one endpoint

Can Strix Halo run Kimi K3?

Next steps