holo-desktop 0.1.0

Vision-language agent that drives real macOS, Linux, and Windows apps. Powered by Holo3.

Holo Desktop

Tell your computer what to do. Holo gets it done. Open-source agent powered by Holo3, H Company's open-weight vision-language model. Use the hosted API, or run everything directly on your computer for full privacy.

Quickstart

pip install holo-desktop
holo run "Catch me up on Slack"

On first run:

1. cua-driver auto-installs (macOS, Linux, Windows).
2. Your browser opens to sign in at portal.hcompany.ai. Skip with --base-url for a local model.
3. macOS only: grant your terminal Accessibility and Screen Recording in System Settings → Privacy & Security.

Run from the terminal

holo run "Open Safari and go to hcompany.ai"

Pin Holo to a specific window with --app:

holo run "Reply to the highlighted Slack thread" --app Slack

See holo run --help for all flags.

Use from Python

from holo_desktop import Holo

answer = Holo().run("Tell me how many unread emails I have")
print(answer)

The Holo class supports pause, resume, stop, and mid-run send for interactive embedding.

Models

Holo defaults to the H Company Models API. Your first holo run opens your browser, signs you in at portal.hcompany.ai, and saves a key to ~/.holo/.env. Run holo login to do this ahead of time. Holo3-35B is on the free tier; the 122B requires a paid plan.

To run everything on your own hardware instead, pass --base-url to any OpenAI-compatible server. No holo login needed, and no screenshots, keystrokes, or app content ever leave your machine.

holo run --base-url http://localhost:8000/v1 "Open Safari and go to hcompany.ai"

Holo3-35B-A3B fits comfortably on a recent MacBook Pro or Mac Mini at Q4. NVIDIA's DGX Spark runs both the 35B and 122B at higher precision and gives you the best agent quality on a single box. Multi-GPU rigs and rack servers serve the FP8 stack at full throughput.

vLLM and llama.cpp both work:

vLLM (Holo3-35B-A3B-FP8)

Per-request reasoning_effort is honored via chat_template_kwargs; think tokens are decoded with --reasoning-parser qwen3.

export VLLM_ATTENTION_BACKEND=FLASHINFER
export TORCH_CUDA_ARCH_LIST=12.1a

vllm serve Hcompany/Holo3-35B-A3B-FP8 \
  --host 0.0.0.0 --port 8000 \
  --tensor-parallel-size 1 --gpu-memory-utilization 0.85 \
  --max-model-len 65537 --max-num-batched-tokens 8192 --max-num-seqs 1 \
  --kv-cache-dtype fp8 --attention-backend flashinfer --enable-prefix-caching \
  --load-format fastsafetensors \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 \
  --chat-template-content-format openai \
  --limit-mm-per-prompt '{"image": 1}' \
  --mm-processor-cache-gb 4 --mm-processor-cache-type shm \
  --trust-remote-code

llama.cpp (Holo3-35B-A3B GGUF)

Quants by mradermacher/Holo3-35B-A3B-GGUF (community).

Reasoning behavior is fixed at server launch (--reasoning auto separates <think> from content). chat_template_kwargs is silently ignored, so per-request reasoning_effort falls back to logit-bias steering on the </think> token.

llama-server -hf mradermacher/Holo3-35B-A3B-GGUF:Q4_K_M \
  --host 0.0.0.0 --port 8000 \
  --jinja --reasoning auto \
  -c 65536 -ngl 99 \
  --chat-template-kwargs '{"enable_thinking": true}'

Use inside another agent

Holo runs as a sub-agent of Claude Code, Cursor, Codex, and other MCP / ACP hosts. When your main agent needs to read a screen or click through an app, it delegates to Holo and gets the answer back.

One command wires Holo into every supported host on your machine:

holo install               # everything detected
holo install cursor        # one host
holo install list          # see what's available

Each host gets the MCP server in its config, plus a Skill (where supported) that teaches the parent when to delegate to Holo.

id	host	skill auto-load
claude-code	Claude Code	~/.claude/skills/
claude-desktop	Claude Desktop	—
codex	Codex	~/.agents/skills/
copilot	GitHub Copilot CLI	—
cursor	Cursor	—
gemini	Gemini CLI	—
hermes	Hermes	—
openclaw	OpenClaw	~/.openclaw/skills/
opencode	OpenCode	~/.config/opencode/skills/

ACP

holo acp runs Holo as an ACP sub-agent over stdio.

Hermes (NousResearch):

delegate_task(acp_command="holo acp", task="Open Authy and grab my AWS 2FA code")

OpenClaw — ~/.openclaw/openclaw.json:

{ "runtimes": { "holo": { "runtime": "acp-standard", "command": "holo", "args": ["acp"] } } }

Zed — ~/.config/zed/settings.json:

{ "agent_servers": { "Holo Desktop": { "command": "holo", "args": ["acp"] } } }

JetBrains, Neovim (avante.nvim, CodeCompanion.nvim): point any ACP client at holo acp.

Develop

git clone https://github.com/hcompai/holo-desktop && cd holo-desktop
make setup

See CONTRIBUTING.md and RELEASING.md.

Resources

• Models: Holo3-35B-A3B · Holo3-122B-A10B
• Docs: Quickstart · Models API
• H Company