DAAK Duffy

You give Duffy a task from the web dashboard, the command line, a chat bridge or the macOS app. Duffy runs it through the AI command-line agent you already use (Claude Code claude -p, Antigravity agy, Ollama ollama run, or any other CLI), falls back to the next model when one runs out of quota, and reports back.

When the agent wants to do something you should sign off on, such as sending an email, pushing code or deleting files, Duffy stops, shows you the exact command, and runs it only after you press Approve.

One Python file, standard library only. Happy on an old laptop or a small home server that never sleeps.

Quick start

git clone https://github.com/benfirad/daak-duffy.git && cd daak-duffy
python3 duffy.py daemon          # worker + dashboard on http://127.0.0.1:8899
python3 duffy.py add "Summarise yesterday's server logs and list anything unusual"

On a Mac, add the menu bar app with cd macos && ./build.command.

Why Duffy

Most agent tools either answer while you watch, or run unattended and hope for the best. Duffy sits in between: autonomous for routine work, human-in-the-loop for anything with consequences.

	Chat with an AI	Agent running unattended	Duffy
Works while you're away	✗	✓	✓
Queue, history and logs that survive restarts	✗	sometimes	✓
Asks before risky actions, shows the exact command	n/a	✗	✓
Switches model when one hits its quota	✗	rarely	✓
Uses the CLI and subscription you already have	✗	varies	✓
Runs on your own machine, no account, no cloud	✗	varies	✓

Good for: inbox and newsletter summaries, research with sources, checking why a backup or cron job failed, keeping a home server healthy, small code chores in your repos, drafting replies that you approve before they go out.

What you get

Task queue	SQLite queue with a full log per task. Survives restarts.
Approval gate	The agent ends its answer with [APPROVAL_REQUIRED: what and why | CMD: command]. Nothing runs until a human approves.
Model fallback	Give it a list of models; quota or rate-limit answers move it to the next one.
Dashboard	Queue, results, approve / reject / retry, light and dark mode.
Chat agents	Optional personas you can talk to; each reply runs through the same agent command.
Notifications	Any webhook that accepts {"chatId", "message"} — WhatsApp or Telegram bridges fit.
Office panel (optional)	office.py: approvals, health checks, launchd services and log tails on one screen.
ai command (optional)	Asks Claude; when Claude's usage limit is hit it switches to Antigravity, then to Duffy, until the limit resets.
macOS app (optional)	Native menu bar app: chat with agents, queue tasks, approve commands, watch services. See below.

flowchart LR
    A["You: dashboard / CLI / chat"] --> Q[("Task queue")]
    Q --> W["Duffy worker"]
    W --> M{"Agent CLI<br/>model 1 → 2 → 3"}
    M -->|report| D["Done + notify"]
    M -->|APPROVAL_REQUIRED| G["Waiting for you"]
    G -->|Approve| R["Run the exact command"] --> D
    G -->|Reject| X["Rejected"]

The macOS app

A native SwiftUI menu bar app for people who would rather not keep a browser tab open. English or Turkish, follows your system language.

Build it with cd macos && ./build.command (macOS 13+, Xcode Command Line Tools, no Xcode project). Details in macos/README.md.

Usage

Open http://127.0.0.1:8899 and queue a task, or from another terminal:

python3 duffy.py add "Summarise yesterday's server logs and list anything unusual"
python3 duffy.py list
python3 duffy.py show 1

By default Duffy calls agy --model <model> -p <prompt>. To use something else, set agent_command in ~/.duffy/config.json (see below).

Configuration

Copy config.example.json to ~/.duffy/config.json. Every key is optional.

Key	Default	Meaning
agent_command	["agy", "--model", "{model}", "-p", "{prompt}"]	Command run per attempt. {model}, {prompt} and {workspace} are filled in.
models	["default"]	Tried in order until one answers.
quota_markers	quota / rate limit texts	Short answers containing these count as "try the next model".
bind_hosts	["127.0.0.1"]	Addresses the dashboard listens on. Add a VPN address (e.g. your Tailscale IP) to reach it from other devices.
allowed_clients	[] (anyone who can connect)	IPs or CIDR ranges allowed to use the dashboard, e.g. ["100.64.0.0/10"].
notify.webhook_url, notify.chat_id	empty	Where to send "done" and "needs approval" messages.
agents	one duffy persona	Chat personas: {"id": {"name": "...", "prompt": "..."}}.

Examples for agent_command:

["claude", "-p", "{prompt}", "--model", "{model}"]   // Claude Code
["ollama", "run", "{model}", "{prompt}"]               // local models
["agy", "--model", "{model}", "-p", "{prompt}"]        // Antigravity CLI

DUFFY_HOME, DUFFY_CONFIG, DUFFY_PORT and DUFFY_BIND override the file.

Commands

Command
duffy.py daemon	Worker loop + dashboard
duffy.py add "task" [-c category]	Queue a task
duffy.py list / show <id>	Recent tasks / one task with its log
duffy.py approve <id> / reject <id> / retry <id>	Decide on a gated task, or re-queue
duffy.py step	Process one task and exit (handy for cron)

The HTTP API mirrors these: GET /api/status, GET /api/logs?id=, and POST /api/add, /api/approve, /api/reject, /api/retry, /api/chat with a JSON body.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗