Local AI setup

This is the local setup path within Agents. Use it for models running on your computer; the main guide covers coding agents and provider sign-in.

The open backend can use a local chat or coding model through OpenCode. The guided Local AI, chat, dictation, and speech screens below belong to the full Desktop client, not the public Agent + Buzz browser. OpenCode configuration and backend integrations remain available to other clients. Ollama runs the model on your computer. Most local models connect through OpenCode. The guided MiMo setup uses a lightweight chat connection directly to Ollama. An optional local voice service can read chat text aloud. Local image tools remain separate experiments. The setup helper can point you toward a useful first step.

Optional setup helper

Find your starting path

Five quick answers create a setup checklist. Your answers stay on this page.

Question 1 of 5

Do you have Glowbom OSS and OpenCode installed?

Yes, both are readyI need to install themI'm not sure

BackNext

Connect a downloaded model in Glowbom

Open Settings → Tools → Local AI, or choose Set up local AI at the bottom of the chat model menu. Keep Ollama running, choose a downloaded model, and click Connect & test. Finish active chats and builds first because connecting a new model reloads Glowbom's OpenCode server.

Glowbom saves this connection separately from your existing OpenCode providers. It checks a short reply with tools disabled and no project files. MiMo chats directly with Ollama to keep coding-agent instructions out of its context. Glowbom sends recent messages, attached images, and a short project summary, not the full prototype or code. It requests an 8,192-token context for these chats without changing Ollama's global settings. Older messages stay visible but may be left out of the model request. Choose another model to discuss project code in detail. Other connected local models use OpenCode. Ready for Chat means that test succeeded. Choose the model under Ollama (this computer) in the model menu to start talking. MiMo remains Chat only.

This flow connects models already downloaded in Ollama. It does not install Ollama or download model files. If you use an external OpenCode server, use the manual configuration below on that server. Glowbom's saved connections live in ~/.glowbom/local-ai.json and apply to the OpenCode server it starts, not to separate OpenCode sessions you launch yourself.

Try MiMo 9B with Ollama

This example starts on a Mac. For other computers, check the linked Ollama and OpenCode guides. A local model needs enough memory and may be slower or less reliable than a hosted model for some tasks.

The linked MiMo V2.6 9B Ollama model is a community upload of Xiaomi's 9B distillation. The Ollama page lists a 6.5 GB download. That is the download size, not a promise about memory needed while it runs. Check the community upload and the upstream model page before downloading. Xiaomi's upstream card marks the model MIT licensed; the community upload has its own page and files to review.

1. Install Ollama and open its app. Set up Glowbom Desktop and OpenCode for the chat steps below. OSS users can use the manual OpenCode configuration with a suitable coding model.

2. Run the model in a terminal. The first run downloads its files:

   ollama run maternion/mimo-v2.6:9b

   Ask it a short question to confirm it responds. Type /bye to leave the terminal chat. Keep the Ollama app running for the next steps.

3. Use Connect & test in Glowbom as described above. For manual setup, add the following Ollama provider to your existing OpenCode configuration. OpenCode's Ollama setup guide explains where opencode.json lives. If you already use other providers, merge this ollama entry into your existing provider object rather than replacing the file:

   {
     "$schema": "https://opencode.ai/config.json",
     "provider": {
       "ollama": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "Ollama (local)",
         "options": {
           "baseURL": "http://localhost:11434/v1"
         },
         "models": {
           "maternion/mimo-v2.6:9b": {
             "name": "MiMo V2.6 9B",
             "attachment": true,
             "modalities": {
               "input": ["text", "image"],
               "output": ["text"]
             }
           }
         }
       }
     }
   }

   For a custom Ollama model that can read images, attachment: true and modalities.input let OpenCode pass image attachments to it. The MiMo tag above lists vision on its Ollama page. For another model, check that the exact tag supports images before using these fields. They cannot add image support to a text-only model. After editing the configuration, restart OpenCode. In Glowbom's model menu, click the Refresh models icon. If the local OpenCode server still uses the old configuration, restart the Glowbom stack and refresh the model list again.

4. In Desktop, turn off Build on send, then choose the connected MiMo model from the chat model menu. Ask a short question or attach a picture. MiMo is marked Chat only because its Build responses through OpenCode can be incomplete. Choose a different model for Draw or Build. The minimal OSS interface has no general Chat screen; use a suitable coding model for its Build workflow.

5. If you choose another local model for Build, check Ollama's context length guide. OpenCode's coding workflow needs a longer context, which also uses more memory. Review and test the result before using it in a real project.

If the model does not appear after refreshing, check that Ollama is running, the model name in opencode.json matches ollama list, and opencode models ollama lists the configured model. ollama ps shows the loaded model and its allocated context after a request. The Ollama model page reports a large maximum context, but the memory your computer can actually allocate is the useful limit.

Local voice

Glowbom can read chat aloud through either KittenTTS Mini or VoiceStudio. They are separate programs. Both use 127.0.0.1:3900, so stop one before starting the other. You can install both, but Glowbom uses one running service at a time. Choose a setup path below, then finish in Settings → Voice → Local voice.

KittenTTS Mini

This Glowbom OSS companion offers eight preset English voices and does not clone voices. It needs the Glowbom OSS source folder and Python 3.8 or later. On macOS or Linux:

1. In a terminal, open the source folder that contains extras/kitten-tts/server.py, then install the voice package:

   cd extras/kitten-tts
   python3 -m venv .venv
   .venv/bin/python -m pip install https://github.com/KittenML/KittenTTS/releases/download/0.8.1/kittentts-0.8.1-py3-none-any.whl

2. Start the companion from that same folder:

   .venv/bin/python server.py

   The first start downloads the Mini model. Leave the terminal open until you finish using local voice. Wait for Local voice ready before refreshing voices in Glowbom. Press Ctrl+C to stop it.

VoiceStudio

VoiceStudio is a separately installed desktop app with its own speech engines. It does not come with the KittenTTS Mini companion. KittenTTS can also run inside VoiceStudio; that uses VoiceStudio's service, not the Glowbom companion server.

1. Get the current Electron desktop app from VoiceStudio's releases and follow its installation guide for your platform. Its current local backend does not support Intel Macs; see the macOS guide.
2. Stop KittenTTS Mini if it is running, then open VoiceStudio. Complete its local runtime setup. For a first connection test, use its default speech engine and download the model when prompted. You can choose another engine in Model Catalogue → TTS; compare the engines and check their model terms before business use. Generate a short sample in VoiceStudio. Its desktop app manages the local backend on port 3900. Keep VoiceStudio open while using it in Glowbom.

Finish in Glowbom

Open Settings → Voice, choose Local voice, then open the setup guide for the service you started. Click Refresh voices, select an available voice, and play a sample. The guide selection does not switch the running service. If no voices appear, check that your service is running on port 3900 and that the other service is stopped. No ElevenLabs key is needed for local voice. Glowbom sends speech requests to the local service; check the chosen engine's behavior and terms before business use.

KittenTTS Mini's weights and library are marked Apache-2.0. VoiceStudio has a separate application license, and its engines and model weights can have different terms.

Explore image tools

The chat and build model menu also includes other connected AI agents. Local image generation from the following tools is a separate experiment. You can add a file you create to your project yourself.

Tool	What to try	Terms to check
Bonsai Image 4B MLX	Generate an image on Apple Silicon using its upstream instructions, then add the file to a Glowbom project.	Its model card marks it Apache-2.0. Check the exact version and its dependencies before business use.
Qwen-Image-2.1	Explore image generation and editing for research or evaluation.	Its current research license requires separate permission for commercial use of the model. Do not use it to make assets for a commercial project without that permission.

Check terms before publishing

The app that runs a model, the model weights, and any other downloaded assets can have different licenses. Read the terms for the exact version you install, especially before business use. A model's license can restrict how you use the model without stating a blanket rule for every image it produces. Review generated work for other people's rights before publishing it.

These links and license summaries were checked on September 26, 2026. Terms and model files can change, so use the linked source pages for the current terms.

Dictate a message locally

Glowbom can turn microphone audio into an editable message using whisper.cpp. This follows the local Whisper approach used by Omarchy's dictation. Voice playback is a separate setting.

Open Local AI → Dictation to check the engine and model. On a Mac with Homebrew, run:

brew install whisper.cpp
mkdir -p ~/.glowbom/models/whisper
curl --fail --location https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin -o ~/.glowbom/models/whisper/ggml-base.en.bin

The English model downloads about 148 MB. Choose Check again, close settings, and click the microphone beside Send. Allow microphone access when asked. Click Done to transcribe, or Cancel to discard the recording. Recording stops automatically after one minute. Review the draft before sending it.

For Linux or a custom installation, follow the whisper.cpp installation guide. Set GLOWBOM_WHISPER_BIN to the executable path and GLOWBOM_WHISPER_MODEL to a compatible GGML Whisper model path before starting Glowbom. Dictation currently uses English transcription. Ollama and a cloud API key are not required.

Audio stays on the computer running Glowbom's backend. Temporary audio and transcript files are removed after transcription, including failures. Dictated text becomes part of the draft and is sent to the selected chat model only when you choose Send. If access is denied, enable the microphone in your browser or system settings and try again. Browsers need localhost or HTTPS for recording.