Install

Three paths. All three give you do and doer on your $PATH.

pipx (recommended)pipcurl (binary, coming soon)

Isolated, auto-updatable, no venv shenanigans.

pipx install doer-cli

Upgrade later:

pipx upgrade doer-cli

Any venv. Global if you insist.

pip install doer-cli

Prebuilt binaries will land on GitHub Releases shortly. For now, use pipx.

what to expect

# once released:
curl -sSL https://github.com/cagataycali/doer-cli/releases/latest/download/install.sh | sh

Installs to ~/.local/bin. No Python required.

optional extras

pip install 'doer-cli[mlx]'   # local text inference + LoRA training (Apple Silicon)
pip install 'doer-cli[vlm]'   # vision/audio/video + VLM LoRA (Apple Silicon)
pip install 'doer-cli[hf]'    # huggingface dataset upload for your training corpus
pip install 'doer-cli[all]'   # everything above

Each extra is opt-in. The default install stays lean.

two binaries, one brain

do   "quick question"     # short form — shell loves brevity
doer "longer query"       # long form — when clarity wins

Both resolve to the same entry point.

requirements

requirement	notes
Python 3.10+	only for pip/pipx paths
Ollama or AWS creds	one of the two — doer auto-detects
A model	ollama pull qwen3:1.7b or Bedrock Claude access or MLX base (auto-download)
Apple Silicon (for MLX only)	M1/M2/M3/M4 Mac — opt-in [mlx] extra
Apple Silicon (for multimodal)	[vlm] extra — vision/audio/video via mlx-vlm

pick a backend

doer auto-detects at runtime:

1. --img/--audio/--video flag present and mlx-vlm installed → mlx-vlm (routes to vision/audio/omni model)
2. AWS creds present (AWS_BEARER_TOKEN_BEDROCK / STS / SSO) → bedrock
3. Apple Silicon and strands_mlx installed → mlx
4. Fallback → ollama

Override with DOER_PROVIDER=bedrock|mlx|mlx-vlm|ollama.

bedrock (cloud, default)

Frontier Claude models, zero local compute. Requires AWS credentials and Bedrock model access.

# simplest: bearer token (get one from the Bedrock console)
export AWS_BEARER_TOKEN_BEDROCK=abc...

# or standard AWS creds
export AWS_PROFILE=my-profile

do "reply: ok"
# → ok

Defaults (all overridable via env):

setting	value
model	global.anthropic.claude-opus-4-7
region	us-west-2
context	1M tokens (via context-1m-2025-08-07 beta, auto)
output	128k tokens (Opus 4.7 native max)

Pin a different model:

DOER_BEDROCK_MODEL=us.anthropic.claude-sonnet-4-20250514-v1:0 do "..."
DOER_BEDROCK_MODEL=global.anthropic.claude-opus-4-6  do "..."

ollama (local, private)

No API keys. Runs on your laptop. Great for offline and sensitive work.

# macOS
brew install ollama
ollama serve &
ollama pull qwen3:1.7b

# linux
curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen3:1.7b

# force ollama even when AWS creds exist
DOER_PROVIDER=ollama do "reply: ok"
# → ok

Switch ollama model:

DOER_MODEL=llama3.2:3b do "explain this file"
DOER_MODEL=qwen3:4b    do "write a sql query"

Persist your choice:

echo 'export DOER_PROVIDER=ollama' >> ~/.zshrc
echo 'export DOER_MODEL=qwen3:4b'  >> ~/.zshrc

mlx (Apple Silicon, opt-in)

For on-device inference with a LoRA adapter you trained yourself. Requires Apple Silicon (M-series).

pip install 'doer-cli[mlx]'        # adds strands-mlx + mlx-lm (~500MB)

# first run downloads the base model from HF
DOER_PROVIDER=mlx do "reply: ok"
# → ok

# hot-swap your trained self (see: Train on yourself)
DOER_PROVIDER=mlx DOER_ADAPTER=~/.doer_adapter do "..."

Change base model:

DOER_MLX_MODEL=mlx-community/Qwen3-4B-4bit do "..."

See Train on yourself for the full collect → train → swap loop.

mlx-vlm (multimodal on Apple Silicon, opt-in)

Vision, audio, and video via mlx-vlm — routed automatically when you pass --img, --audio, or --video flags.

pip install 'doer-cli[vlm]'        # adds mlx-vlm + datasets

do --img screenshot.png  "what's in this UI?"
do --audio meeting.wav   "transcribe + action items"
do --video clip.mp4      "what's happening here?"
do --img a.png --audio b.wav "..."   # omni model (auto-picked)

Models are auto-selected based on the modality mix; override with env vars:

modality	default model	env override
image only	mlx-community/Qwen2.5-VL-3B-Instruct-4bit	DOER_MLX_VLM_MODEL
audio only	mlx-community/gemma-3n-E2B-it-4bit	DOER_MLX_AUDIO_MODEL
image + audio	mlx-community/Qwen3-Omni-30B-A3B-Instruct-4bit	DOER_MLX_OMNI_MODEL

Train a VLM LoRA on your own multimodal corpus:

do --train-vlm 300                             # → ~/.doer_vlm_adapter
DOER_PROVIDER=mlx-vlm DOER_VLM_ADAPTER=~/.doer_vlm_adapter do --img x.png "..."

troubleshooting

connection refused (Ollama)

Ollama isn't running. ollama serve in another terminal.

model not found (Ollama)

Pull it first: ollama pull qwen3:1.7b

AccessDeniedException / ValidationException (Bedrock)

Your AWS creds don't have access to the default model. Either request access in the Bedrock console, or switch: DOER_BEDROCK_MODEL=us.anthropic.claude-sonnet-4-20250514-v1:0 do "..."

invalid beta flag (Bedrock)

A model rejected a beta header you set. Clear it: DOER_ANTHROPIC_BETA="" do "..."

temperature/top_p 400 error (Bedrock)

Opus 4.7+ rejects any non-default sampling values. Unset DOER_TEMPERATURE / DOER_TOP_P, or pin an older model.

slow first call

Ollama loads the model into RAM on first call. Subsequent calls are fast. Ambient tip: ollama run qwen3:1.7b "" & to warm it in the background.

command not found: do

pipx installed but ~/.local/bin isn't on $PATH. Run pipx ensurepath and restart your shell.

uninstall

pipx uninstall doer-cli     # pipx users
pip  uninstall doer-cli     # pip users
rm ~/.local/bin/do ~/.local/bin/doer    # binary users

That's it. No config dirs. No leftovers. Nothing to clean up.