doer-cli 0.8.0

doer — one-file pipe-native agent. strands-agents + ollama only.

DOER

stdin → agent → stdout

A Unix citizen that thinks. In text, images, audio, and video.

📖 Full documentation →

---

install

# pipx — isolated, auto-updatable (recommended)
pipx install doer-cli

# pip — any venv
pip install doer-cli

# optional extras
pip install 'doer-cli[mlx]'   # local inference + LoRA training (Apple Silicon)
pip install 'doer-cli[vlm]'   # vision/audio/video + VLM LoRA
pip install 'doer-cli[hf]'    # huggingface dataset upload
pip install 'doer-cli[gr00t]' # Isaac GR00T policy-server client (ZMQ)
pip install 'doer-cli[all]'   # everything

Two binaries land on $PATH: do (short) and doer (long).

do is a shell keyword in bash/zsh loops — the binary still works because of the argument (do "hi" is unambiguous), but if tab-completion misbehaves, alias it:

echo 'alias do="doer"' >> ~/.zshrc

run

# text
do "find files larger than 100MB"

# pipe
cat error.log | do "what broke"
git log -20   | do "write release notes"
curl -s api.io | do "summarize" | tee out.md

# multimodal (auto-routes to mlx-vlm on Apple Silicon)
do --img screenshot.png "what's in this UI?"
do --audio meeting.wav   "transcribe and bullet the action items"
do --video clip.mp4      "what's happening here?"
do --img a.png --audio b.wav "..."   # omni model (auto-picked)

what it is

Agent(
    model=<auto: bedrock | ollama | mlx | mlx-vlm>,
    tools=[shell] + hot_reload("./tools"),
    system_prompt=SOUL.md + AGENTS.md + ~/.doer_history
                + ~/.bash_history + ~/.zsh_history + own_source,
)(stdin + argv + images + audio + video)

One file. doer/__init__.py — ~730 lines. Reads your shell like a person reads a room. Trains on its own transcripts. Swaps its own brain.

context it sees every call

source	what
SOUL.md (cwd)	who it is in this project
AGENTS.md (cwd)	rules for this project
~/.doer_history	last N Q/A (DOER_HISTORY=10)
~/.bash_history + ~/.zsh_history	last N commands (DOER_SHELL_HISTORY=20)
./tools/*.py	hot-reloaded @tool functions
own source	full self-awareness
--img / --audio / --video	raw media sent to a VLM (routed automatically)

No database. No config file. The filesystem is the memory.

providers

Auto-picked from what's on your machine. Override with DOER_PROVIDER.

provider	when	model default
bedrock	AWS creds present (AWS_BEARER_TOKEN_BEDROCK / STS / SSO)	global.anthropic.claude-opus-4-7 (1M ctx)
mlx-vlm	Apple Silicon + [vlm] extra + --img/--audio/--video	Qwen2.5-VL-3B / gemma-3n / Qwen3-Omni
mlx	Apple Silicon + [mlx] extra (text-only, trained adapter)	mlx-community/Qwen3-1.7B-4bit
ollama	fallback — local, private, no keys	qwen3:1.7b

# force a provider
DOER_PROVIDER=ollama do "quick ping"
DOER_PROVIDER=mlx DOER_ADAPTER=~/.doer_adapter do "use my trained self"

env knobs

# provider selection
DOER_PROVIDER=                   # "" (auto) | bedrock | ollama | mlx | mlx-vlm

# bedrock (defaults tuned for Claude Opus 4.7)
DOER_BEDROCK_MODEL=global.anthropic.claude-opus-4-7
DOER_BEDROCK_REGION=us-west-2
DOER_MAX_TOKENS=128000           # Opus 4.7 native max
DOER_TEMPERATURE=                # unset on Opus 4.7+ (returns 400 otherwise)
DOER_TOP_P=                      # unset on Opus 4.7+
DOER_CACHE_PROMPT=               # "1" / "true" → prompt caching
DOER_ANTHROPIC_BETA=context-1m-2025-08-07   # csv; auto on Claude — "" to disable
DOER_ADDITIONAL_REQUEST_FIELDS=  # raw JSON escape hatch
DOER_BEDROCK_GUARDRAIL_ID=       # optional Bedrock guardrail
DOER_BEDROCK_GUARDRAIL_VERSION=

# ollama
DOER_MODEL=qwen3:1.7b
OLLAMA_HOST=http://localhost:11434

# mlx (Apple Silicon)
DOER_MLX_MODEL=mlx-community/Qwen3-1.7B-4bit
DOER_ADAPTER=                    # path to trained text LoRA
DOER_MLX_VLM_MODEL=mlx-community/Qwen2.5-VL-3B-Instruct-4bit
DOER_MLX_AUDIO_MODEL=mlx-community/gemma-3n-E2B-it-4bit
DOER_MLX_OMNI_MODEL=mlx-community/Qwen3-Omni-30B-A3B-Instruct-4bit
DOER_VLM_ADAPTER=                # path to trained VLM LoRA

# context
DOER_HISTORY=10                  # Q/A rows in prompt
DOER_SHELL_HISTORY=20            # shell rows in prompt
DOER_DEBUG=                      # "1" → verbose errors
DOER_MAX_SEQ_LEN=16384           # LoRA training max seq length

# huggingface upload
DOER_HF_REPO=<user>/doer-training   # override target
HF_TOKEN=                           # or `huggingface-cli login`

extend in 60 seconds

# ./tools/weather.py
from strands import tool
import urllib.request

@tool
def weather(city: str) -> str:
    """Weather for a city."""
    return urllib.request.urlopen(f"https://wttr.in/{city}?format=3").read().decode()

Next call: do "istanbul weather?" — hot-reloaded, no restart.

the loop: collect → train → swap

doer closes its own training loop. Every call appends a dense, self-contained record to ~/.doer_training.jsonl — full system prompt, all messages, tool specs, native <tool_call> tokens preserved.

# 1. collect (automatic — just use doer)
do "fix this stacktrace" < err.log
do --img ui.png "label the bugs in this screenshot"
# ... 100+ real turns across text/image/audio/video

# 2. inspect
do --train-status
# → 127 turns | 2453.1KB | sha256:250c406b | ~/.doer_training.jsonl
#     text:102  image:18  audio:3  video:4
#     hf:    cagataydev/doer-training | in sync

# 3. train — in-process LoRA (no trainer indirection, ~50 lines calling mlx_lm.tuner)
do --train 200                # text LoRA     → ~/.doer_adapter
do --train-vlm 300            # vision LoRA   → ~/.doer_vlm_adapter

# 4. use your trained self
DOER_PROVIDER=mlx     DOER_ADAPTER=~/.doer_adapter         do "fix this" < err.log
DOER_PROVIDER=mlx-vlm DOER_VLM_ADAPTER=~/.doer_vlm_adapter do --img x.png "what's this?"

Training preserves native tool-call tokens via the tokenizer's chat template — your adapter learns real tool-use, not string mimicry.

train in the cloud (HuggingFace Jobs)

Laptop LoRA is great for 500-turn datasets on Qwen3-1.7B. When you want to scale up — bigger models, full fine-tunes, VLM, Omni — burn HF credits instead of your battery.

# one-shot dispatchers (UV scripts, zero setup)
doer --hf-jobs gen                     # Generate dataset (CPU, ~$0.60/500 prompts)
doer --hf-jobs text                    # Qwen3-1.7B LoRA, T4, ~$0.30
doer --hf-jobs vlm                     # Qwen2.5-VL-3B LoRA, A100, ~$5
doer --hf-jobs omni                    # Qwen2.5-Omni-7B, H200, ~$10

# generate dataset from your own prompts (one per line, dedupes against existing)
doer --hf-jobs gen my_prompts.txt --iters 500
doer --hf-jobs gen hf://Anthropic/hh-rlhf:chosen --iters 1000

# override anything via env or flags
MODEL=Qwen/Qwen3-4B FLAVOR=a10g-large doer --hf-jobs text --iters 1000
PROVIDER=ollama CONCURRENCY=16 doer --hf-jobs gen prompts.txt

# monitor
doer --hf-jobs ps
doer --hf-jobs logs <job_id>
doer --hf-jobs hw          # list hardware + cost/hour

Under the hood each dispatcher is one self-contained UV script bundled inside doer/hf_jobs/ — no repo cloning, no Dockerfile. The script pulls cagataydev/doer-training (your dataset), runs SFT LoRA with trl + peft, merges the adapter, and pushes the full merged model to cagataydev/doer-<model-short> automatically.

Validated end-to-end on T4-medium ($0.60/hr):

• 522 turns → 468 train / 53 eval
• Qwen3-1.7B, LoRA r=16 (17.4M params, 1% of base)
• 50 steps, 33 min → eval_loss 0.149, token accuracy 97.6%
• 3.44 GB merged model auto-pushed to private HF repo

Use the trained model anywhere:

DOER_PROVIDER=transformers DOER_MODEL=cagataydev/doer-qwen3-17b do "what is doer"

See the bundled hf_jobs/README.md (also accessible at $(doer --hf-jobs)/README.md after install) for full details, the generator (gen_dataset.py), and the three trainers (train_text_lora.py, train_vlm.py, train_omni.py).

share the dataset (HuggingFace)

pip install 'doer-cli[hf]'

do --upload-hf                       # → <user>/doer-training (private)
do --upload-hf cagataydev/my-data    # custom repo
do --upload-hf-public                # public dataset

Idempotent — one atomic commit per run (train.jsonl + README with schema/stats/sha). Reuses huggingface-cli login or HF_TOKEN. --train-status shows local sha vs last remote commit.

Round-trip anywhere:

hf download cagataydev/doer-training --repo-type dataset --local-dir /tmp/d
cp /tmp/d/data/train.jsonl ~/.doer_training.jsonl
do --train 200

robotics: Isaac GR00T (v0.8.0+)

doer speaks the Isaac GR00T policy-server protocol (ZMQ REQ/REP, msgpack) natively. Use it as the brain that plans robot actions, or as a pipe that converts observation JSON into action JSON — no LLM, no tokens, just stdin → action.

pip install 'doer-cli[gr00t]'   # adds pyzmq + msgpack + numpy + pillow

pipe mode — raw ZMQ client, LLM bypass:

echo '{"state.joint_pos":[0,0,0,0,0,0,0]}' | doer --gr00t "pick up cube"
# → {"action":{"action.joint_pos":[[0.01,...]]}, "info":{"inference_time_ms":42}}

brain mode — LLM calls gr00t_action as a tool:

export DOER_GR00T_HOST=thor.local
do "snap /tmp/cam.jpg with libcamera-still, then tell gr00t to stack red on blue"

auto-spawn the server (on a CUDA box):

doer --gr00t-serve /path/to/gr00t-n1.7-ckpt --embodiment-tag new_embodiment

Helpers: doer --gr00t-ping · doer --gr00t-schema · doer --gr00t-reset. Env knobs: DOER_GR00T_HOST / PORT / EMBODIMENT / TIMEOUT_MS / API_TOKEN. Every gr00t_action call lands in ~/.doer_training.jsonl as a native tool-use record — doer --train learns when to ask the policy server from the log.

philosophy

┌─────┐       ┌──────┐       ┌──────┐
│stdin│──────▶│  do  │──────▶│stdout│
└─────┘       └──────┘       └──────┘

grep with a brain. Chain it. Script it. Cron it.

Read SOUL.md for the manifesto. Read AGENTS.md for the rules.

family

project	size	purpose
doer	~730 LOC	one pipe, one shell, one file, one loop (collect→train→swap)
DevDuck	60+ tools	every protocol, every edge

uninstall

pipx uninstall doer-cli    # or: pip uninstall doer-cli
rm /usr/local/bin/do       # if installed via binary

license

Apache-2.0 · built in New York · 2026

---

do one thing and do it well — Doug McIlroy, 1978