Connecting apps

Connecting apps

TokenStation5 speaks the OpenAI API on port 8081, over plain HTTP. Any app with an "OpenAI compatible" provider works with three settings:

Setting	Value
Base URL	http://<the PS5's address>:8081/v1 (http, not https, ending in /v1)
Model	the id that http://<the PS5's address>:8081/v1/models shows, such as Qwen2.5-Coder-7B-Instruct-Q4_K_M (the server answers with the model it has loaded whatever id is sent)
API key	one of your keys (Settings > Apps > New key). With "Require a key" off any text works, and a real key still shows the app by name on Home

Settings > Apps > Instructions > "Set up an app" shows these on the TV, filled in. A DHCP reservation in your router keeps the address from changing.

/v1/models also gives the loaded model's context_length: tell your app that number as the context window, so it trims the conversation before the PS5 has to refuse it. Coding tools send a long first prompt (10,000 to 15,000 tokens of instructions and tool descriptions), so serve with as much context as fits (the default). It lists the model's sampling settings under settings too: the temperature, top_p and top_k a request gets when it sends none, which for a checked model are the ones that did best on the PS5. Leave them out of your app's config to get them.

Endpoints: /v1/chat/completions (streaming, tool calling), /v1/completions, /v1/models.

API keys

Settings > Apps > New key: name it after the app, and the key shows once on its row, like ts5-k7qp2-wx9md-3hf8r-tz6na (no i, l, o or u, so it reads off the TV without mix-ups). Type it into the app then; the PS5 keeps only its SHA-256. Cross turns a key off and on, square twice removes it. With Require a key on, a request without a key that's on gets HTTP 401; the Chat page on the TV never needs one. ts5 serve on a computer takes one key from TS5_API_KEY.

Kilo Code

Kilo Code reads custom providers from ~/.config/kilo/kilo.jsonc:

{
  "provider": {
    "tokenstation5": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "TokenStation5",
      "options": { "baseURL": "http://192.168.1.50:8081/v1", "apiKey": "ts5-k7qp2-wx9md-3hf8r-tz6na" },
      "models": {
        "Qwen2.5-Coder-7B-Instruct-Q4_K_M": {
          "name": "Qwen 2.5 Coder 7B",
          "tool_call": true,
          "limit": { "context": 32768, "output": 8192 }
        }
      }
    }
  }
}

Then pick tokenstation5/Qwen2.5-Coder-7B-Instruct-Q4_K_M as the model. Put the context_length from /v1/models in limit.context. Leave reasoning out for the Llama and Qwen 2.5 models; only Qwen 3 thinks.

Continue, Cline, Roo Code, Open WebUI, aider, opencode and others

Choose the "OpenAI compatible" provider (or "OpenAI" with a custom base URL) and fill in the three settings. Where the app asks for a context window, give it the context_length; for max output tokens, 4,096 to 8,192 is plenty. Native tool calling works with every checked model.

Which model for what

• Agent work (editing files, then running them): Qwen 3 8B. On the PS5 it took every step of a write-then-run task; Qwen 2.5 Coder 7B wrote the file and stopped in 4 of 4 runs.
• Qwen 3 thinks before it answers; the thinking comes back apart from the answer as reasoning_content. Skip it with "chat_template_kwargs": {"enable_thinking": false} or /no_think at the end of the message. A reply capped under 1,024 tokens never thinks, so a request for a short title still gets an answer.
• Llama 3.1 8B and 3.2 3B sometimes call a tool the question doesn't need (asked for a poem about rain with only a weather tool, Llama 3.1 8B asks for the weather). For plain chat with a Llama, leave the tools out, or use a Qwen model.
• Mistral Nemo 12B takes tools in its own format and passed every tool check on the PS5, with about 15k of context beside its weights.
• Gemma 3 12B takes tools in Qwen 3's format inside Gemma's own turns; 59 of 60 checks passed over three runs on the PS5, with about 29,700 tokens of context.

When it doesn't connect

• The Logs page has a line for each request. A request that never shows there never reached the PS5: check the address, and that the app uses http://. An app that tries https:// is named in the log.
• "maximum context length" errors: the conversation is longer than the model's context. Serve with as much context as fits, or start a new task in the app.
• Away from home, see Tailscale. Don't forward port 8081 on your router.