Connect your own AI
Two ways in: paste a key from OpenAI, Google Gemini or OpenRouter, or point Speakybara at a model you already run yourself with Ollama, LM Studio, llama.cpp or Unsloth/vLLM.
Explain with AI, Summarize, Translate, Improve writing, Step-by-step plan, the optional clean-up, word meanings and every action you write yourself need a model to talk to. Speakybara runs none of its own: you connect one under Settings › AI for tasks. That is a key of your own from an online service, or a model running on your own PC.
You need Speakybara installed, and either an account with OpenAI, Google (AI Studio) or OpenRouter to create a key in, or Ollama, LM Studio, llama.cpp or Unsloth/vLLM already running on your PC. Reading aloud needs none of this: the voice is a separate choice, under Settings › Voice.
Connecting a model
Step 1
Open Settings › AI for tasks
The page asks one question: Who helps with your tasks? Off means reading aloud only. Speakybara AI belongs to the Compleet subscription and cannot be chosen yet. For an AI of your own, pick Your own key or Your own model; only the answer you pick shows its fields.
On a Mac the page has three answers, Compleet, Local and API, with a switch Use AI for tasks at the top. Local is a model on your Mac, API a key of your own.
Step 2
With a key: pick the service and paste it
Choose Your own key, pick OpenAI, Google Gemini or OpenRouter as the AI service and paste your key in the Key field. The model fills itself in: gpt-4o-mini for OpenAI, gemini-3.8-flash for Google Gemini, and the same Gemini model for OpenRouter (google/gemini-3.8-flash). A Gemini key comes from Google AI Studio; for OpenRouter it is the same key as for the OpenRouter voices. The key stays in config.json on your own PC and goes only to that service, with each request.
Another service that speaks the OpenAI format works too: choose Own server address and type its address.
Step 3
With a model on your PC: pick your program
Choose Your own model, then Ollama, LM Studio, llama.cpp or Unsloth / vLLM as the program, and the right address fills itself in: port 11434, 1234, 8080 and 8000 respectively. No key, and nothing goes over the internet.
The address is under More options › Server address, for a server that listens on another port.
Step 4
Not sure which model your server has loaded
Press Find models next to the Model field. Speakybara asks the service or your server for its list, with the recommended models at the top, and you pick one; typing a name yourself is fine too.
Step 5
Test it, then run one action
Press Test next to Connection: it says Works, or what went wrong. Then capture a sentence and pick Explain with AI or Summarize under Quick action, in the bar under the frame. An answer read aloud means everything is connected.
Worth knowing
- What is sent is the recognised text, as a chat request to the service you chose; your capture is not sent along. The one exception is off by default: switch on Explanations may look at your screen, under AI for tasks, and asking about a step also sends a picture of your screen to that AI. Speakybara has no server in that path.
- Switch on Clean up text before reading it aloud, on the same page, and every capture goes to that AI first, even with a voice on your PC. It is off by default. On a Mac, Clean up text has a row of its own in Settings, with its own service.
- What the service charges is between you and that service; Speakybara takes no cut and adds nothing. Under More options, Usage this month shows how often each model ran and roughly what it cost.
- A model you run yourself is separate software with its own appetite for memory and often for your graphics card, outside the roughly 55 MB Speakybara uses.
- One action of your own can use a different model from the rest, under that action's Advanced options. On a Mac it can use a different service as well.
If something goes wrong
It stays in config.json, Speakybara's settings file on your own computer, and goes only to the service you chose, as the authorisation on each request. It is not sent anywhere else, and we never see it.
Yes. You use one at a time, but each service keeps its own key and model, so switching back and forth loses nothing. A single action of your own can also be pinned to a different model under its Advanced options.
Check that the program — Ollama, LM Studio, llama.cpp or Unsloth/vLLM — is really running, and that the server address under More options matches its port. The app fills in the usual one, so a server on a different port has to be typed in by hand.
Yes: choose Off at the top of Settings › AI for tasks (on a Mac, switch off Use AI for tasks). Then nothing goes to an AI: no quick actions, no clean-up, no explanations. Your key stays saved for when you switch it back on.