bolotype 0.1.1

System-wide local voice typing with AI-powered editing

BoloType — Type at the speed of sound. Refine with LLMs of your choice.

System-wide voice typing with instant dictation and AI-powered editing.

Platform support: Linux only. Windows and macOS support is coming soon.

---

Features

• Instant dictation — speech inserts at the cursor as soon as a phrase completes. No push-to-talk, no button to hold.
• Bring your own LLM — connect any OpenAI-compatible endpoint: llama.cpp, LMDeploy, vLLM, Ollama, OpenAI, Groq, and more. One setting to change.
• Fully local by default — audio never leaves your machine. LLM calls go only to the endpoint you configure. Nothing is sent anywhere unless you choose a cloud endpoint.
• Private and safe — no cloud account required, no telemetry, no data collection of any kind.
• Voice polish commands — say "polish this paragraph" and the LLM rewrites that span in place. Works on a line, paragraph, selection, or the entire field.
• Precise undo — "undo that" restores the exact pre-polish text, not a generic Ctrl+Z that may affect unrelated changes.
• Two ASR engines — Moonshine is fast, low-latency, and works on CPU. Nemotron 3.5 offers higher accuracy and requires a CUDA GPU.
• System-wide — works in any text field reachable via AT-SPI2: GTK apps, browsers, Electron apps, terminals, and more.

---

Quickstart

Requires Ubuntu/Debian. Moonshine engine used below (CPU-friendly).

python3 -m venv --system-site-packages .venv
source .venv/bin/activate

pip install bolotype
bolotype install moonshine

Edit ~/.bolotype/settings.json and set your LLM endpoint and model name (see Configure), then:

bolotype run --start-active

Say "hello world" — it appears at the cursor. Say "polish this paragraph" — the LLM rewrites it in place.

---

Install

python3 -m venv --system-site-packages .venv
source .venv/bin/activate

pip install bolotype

This installs only the lightweight core (CLI, LLM client, config). Then install your ASR engine:

bolotype install moonshine    # Moonshine — CPU-friendly, low latency (recommended)
bolotype install nemotron     # Nemotron 3.5 — higher accuracy, requires CUDA GPU
bolotype install              # both engines

bolotype install also installs Linux system packages (python3-pyatspi, xdotool, xclip, etc.) via apt and creates ~/.bolotype/ with default config files.

Nemotron note: Nemotron 3.5 has not been tested on CPU. A CUDA-capable GPU is strongly recommended. If you do not have a GPU, use Moonshine.

---

Voice commands

Command	What it polishes
polish this line	The line the caret is on
polish this paragraph	The paragraph the caret is in (blank-line delimited)
polish everything	The entire field
polish the selection	The currently selected text
polish this	Alias for polish everything
undo that	Restore the last polished span

Commands trigger when a completed phrase exactly matches one of the above after punctuation and whitespace normalization.

---

Configure

All settings live in ~/.bolotype/settings.json. Only set the keys you want to override — everything else uses the defaults shown below.

LLM

{
  "llm": {
    "base_url": "http://127.0.0.1:8000/v1",
    "model": "your-model-name",
    "api_key": "",
    "api_key_env": "OPENAI_API_KEY",
    "timeout_seconds": 20,
    "temperature": 0,
    "max_output_tokens": 1000
  }
}

Key	Default	Description
base_url	""	OpenAI-compatible API base URL. E.g. http://127.0.0.1:8000/v1 for a local server or https://api.openai.com/v1 for OpenAI.
model	""	Required. Model name exactly as your server expects it.
api_key	""	API key value. Leave empty if your local server doesn't require one.
api_key_env	"OPENAI_API_KEY"	Name of the environment variable to read the key from. Takes precedence over api_key.
timeout_seconds	20	Seconds to wait for the LLM response before raising an error.
temperature	0	Generation temperature. 0 is deterministic and gives the most consistent edits.
max_output_tokens	1000	Maximum tokens in the polished response.

ASR

{
  "asr": {
    "engine": "moonshine",
    "language": "en",
    "nemotron_model_id": "nvidia/nemotron-3.5-asr-streaming-0.6b",
    "nemotron_lookahead_tokens": 3,
    "nemotron_vad_threshold": 0.01,
    "nemotron_silence_duration_s": 1.25
  }
}

Key	Default	Description
engine	"moonshine"	ASR engine to use. "moonshine" or "nemotron". Can also be set via --asr-engine flag or BOLOTYPE_ASR_ENGINE env var.
language	"en"	Language code. Moonshine accepts "en". Nemotron accepts locale codes such as "en-US", "de-DE", or "auto".
nemotron_model_id	"nvidia/nemotron-3.5-asr-streaming-0.6b"	HuggingFace model ID for Nemotron. Downloaded automatically on first run.
nemotron_lookahead_tokens	3	Controls the streaming chunk size and right-context window. Supported values: 0 (80 ms), 1 (160 ms), 3 (320 ms), 6 (560 ms), 13 (1120 ms). Lower values reduce latency; higher values give the model more context.
nemotron_vad_threshold	0.01	RMS energy level that triggers speech onset detection. Increase if background noise causes false triggers; decrease if the first syllable is being clipped.
nemotron_silence_duration_s	1.25	Seconds of silence after speech ends before the utterance is submitted for transcription. Increase for slower or more deliberate speech.

Input

{
  "input": {
    "backend": "auto",
    "append_space": true
  }
}

Key	Default	Description
backend	"auto"	Keyboard injection backend. "auto" picks xdotool on X11 and ydotool/wtype on Wayland. Set explicitly if auto-detection is wrong.
append_space	true	Append a space after each inserted phrase so the next phrase starts cleanly.

Accessibility

{
  "accessibility": {
    "max_text_characters": 12000
  }
}

Key	Default	Description
max_text_characters	12000	Safety limit. Polish is refused for text spans larger than this. Prevents accidentally sending a very large document to the LLM.

Environment variables

All settings can also be set via environment variables, which take precedence over settings.json.

Variable	Maps to
BOLOTYPE_ASR_ENGINE	asr.engine
OPENAI_BASE_URL	llm.base_url
OPENAI_API_KEY	llm.api_key
VOICE_LLM_MODEL	llm.model
VOICE_LLM_TIMEOUT	llm.timeout_seconds
VOICE_LLM_TEMPERATURE	llm.temperature
VOICE_LLM_MAX_TOKENS	llm.max_output_tokens
BOLOTYPE_CONFIG_DIR	Config directory (default: ~/.bolotype)

Prompt customization

The LLM polish prompt lives at ~/.bolotype/system_prompt.txt and is created by bolotype install. Edit it freely:

bolotype prompt-path          # prints the path
nano ~/.bolotype/system_prompt.txt

---

Run

bolotype run --start-active

All subcommands:

bolotype install moonshine  # install Moonshine ASR + system deps + create config
bolotype install nemotron   # install Nemotron ASR + system deps + create config
bolotype install            # install both ASR engines + system deps + create config
bolotype run                # start the daemon
bolotype toggle             # start/stop listening
bolotype polish             # polish entire focused field
bolotype polish-line        # polish current line
bolotype polish-paragraph   # polish current paragraph
bolotype polish-selection   # polish selected text
bolotype undo               # undo last polish
bolotype status             # show daemon status
bolotype shutdown           # stop the daemon
bolotype config-path        # print ~/.bolotype
bolotype prompt-path        # print ~/.bolotype/system_prompt.txt

For GNOME keyboard shortcuts, use the absolute path to the venv binary so the shortcut works outside a terminal session:

/path/to/bolotype/.venv/bin/bolotype toggle

---

Replacement strategy

Native GTK fields (gedit, GNOME Text Editor, LibreOffice, terminals): AT-SPI EditableText writes the span directly.

Browser/Electron fields (Chrome, Electron apps): AT-SPI Text.setSelection selects the span, the polished text is placed on the clipboard via xclip/wl-copy, and Ctrl+V pastes it.

Fields exposing neither interface (password fields, some embedded widgets) are unsupported.

---

Development

pip install -e ".[dev]"
python -m build
twine check dist/*

---

Known issues

Virtual environment

Ubuntu's pyatspi and gi packages are installed via the system package manager and are not available on PyPI. A normal isolated venv will not see them.

Required:

python3 -m venv --system-site-packages .venv

Conda

Conda environments may be incompatible with Ubuntu's compiled gi bindings due to ABI differences. Recommended pattern: run BoloType in a system-Python venv; run external model servers (LMDeploy, vLLM) in their own conda environment and communicate over HTTP.

Nemotron: GPU required

Nemotron 3.5 is loaded with device_map="auto" and has not been tested on CPU. A CUDA-capable GPU is strongly recommended. If you do not have a GPU, use the Moonshine engine instead.

Nemotron: latency

Nemotron processes audio in streaming chunks but submits complete utterances for transcription. End-to-end latency depends on GPU speed and utterance length. Detailed benchmarking has not been done. Moonshine is lower latency on both CPU and GPU. You are encouraged to tune nemotron_lookahead_tokens and nemotron_silence_duration_s to match your hardware and speaking style.

Browser and Electron accessibility

Chromium-based rich text editors may expose only an object-replacement placeholder (U+FFFC) via AT-SPI. BoloType searches the AT-SPI subtree for the real text node. If none is found, polish is unsupported for that field.

Wayland

Input injection differs between compositors. X11 via xdotool is more reliable. Wayland requires ydotool or wtype.

Rich text

Targeted replacement through accessibility APIs may flatten formatting or operate only on the exposed accessible node.

---

License

MIT © 2026 Muhammad Abdullah Baig