@daomar/agentfleet

AgentFleet

Orchestrate AI coding agents across every machine you own.

Submit one prompt, let every enrolled device run it with its own local agent, and compare results side by side — no server, no SSH mesh, no control plane.

English | 简体中文 · Dashboard · Support the project

Built for agent developers

AgentFleet is designed for developers who work with AI coding agents — Claude Code, GitHub Copilot CLI, Cursor, Aider, or any command-line tool that accepts a prompt and produces code. If you have more than one machine, AgentFleet lets them all work in parallel on the same task and return independent results.

Use it when you want to:

• Compare agents side by side — run the same prompt with Claude Code on one machine and Copilot CLI on another, then compare output quality.
• Parallelize exploratory work — ask multiple machines to inspect different clones, branches, or configurations at the same time.
• Benchmark prompts — evaluate the same prompt across different models, tools, hardware, or repository states.
• Keep data local — tasks and results sync through your OneDrive; AgentFleet runs no hosted backend.
• Work behind firewalls — no inbound ports, tunnels, SSH bastions, queues, or control-plane credentials.

Quick start

One command to start. Run this on every machine you want in the fleet:

npx -y @daomar/agentfleet run

That's it. AgentFleet will:

• Automatically detect your OneDrive and initialize the workspace (first run only).
• Register the agentfleet shortcut command so you can use it directly next time.
• Start watching for tasks.

Then submit a task from any machine:

agentfleet submit \
  --title "Security review" \
  --working-dir /path/to/project \
  --prompt "Review this repository for auth, injection, and secret-handling issues. Return concrete findings and patches."

Every running machine picks up the same task, executes it locally with its configured coding agent, and writes hostname-scoped results back to OneDrive.

Check progress:

agentfleet status

Prefer a browser? Open agentfleet.daomar.dev/web/, sign in with the Microsoft account that owns the OneDrive, submit tasks, monitor nodes, and inspect results.

Agent compatibility

AgentFleet works with any command-line tool that accepts a prompt. The default agent is Claude Code:

Agent	Command template	Notes
Claude Code	claude -p {prompt}	Default, works out of the box
GitHub Copilot CLI	gh copilot suggest "{prompt}"	Requires gh CLI with Copilot extension
Cursor	cursor --prompt "{prompt}"	Requires Cursor CLI
Aider	aider --message "{prompt}"	Requires aider installed
Custom	Any command with {prompt}	Set via defaultAgentCommand in config

Override per-task with --agent:

agentfleet submit --prompt "Fix the login bug" --agent 'aider --message "{prompt}"'

Or change the default globally in ~/.agentfleet/config.json:

{
  "defaultAgent": "claude-code",
  "defaultAgentCommand": "claude -p {prompt}"
}

Concrete workflows

1. Multi-agent code review

Run the same prompt across several machines that each use a different local agent command.

agentfleet submit \
  --title "Compare review agents" \
  --working-dir ~/src/product \
  --prompt "Find the top 10 production risks in this codebase. Group by severity and include file paths."

Use the outputs to compare reasoning quality, false positives, and patch suggestions.

2. Distributed refactor rehearsal

Point multiple machines at separate clones or branches, then let them attempt the same migration independently.

agentfleet submit \
  --title "React hooks migration plan" \
  --working-dir ~/src/frontend \
  --prompt "Plan and implement a safe migration from class components to function components with hooks. Keep changes reviewable."

3. Local benchmark farm for agent prompts

Tune one high-value prompt and run it repeatedly across machines with different models, tools, or repository states.

agentfleet submit \
  --title "Prompt benchmark" \
  --prompt "Summarize build failures, identify the root cause, and propose the smallest fix."

4. Firewalled team or lab environment

Use machines that already sync OneDrive without opening any inbound network path. No extra scheduler needs to be approved, hosted, patched, or monitored.

How it works

Machine A              Machine B              Machine C
agentfleet run         agentfleet run         agentfleet submit
     │                      │                       │
     └──────────────┬───────┴──────────────┬────────┘
                    ▼                      ▼
             OneDrive / AgentFleet shared workspace
                 tasks/  → immutable task JSON
                 output/ → per-machine result files

• agentfleet run detects OneDrive and initializes the shared workspace (auto-init on first run).
• tasks/ and output/ are backed by the synced OneDrive workspace.
• agentfleet submit writes an immutable task file.
• Each agentfleet run process sees new tasks and executes them with the local coding agent.
• Results are written with hostname prefixes, so multiple machines never overwrite each other.

AgentFleet vs. common alternatives

Capability	AgentFleet	SSH scripts	Ansible	Cloud CI
Central server/control plane	No	No, but you manage hosts	Control node	Hosted provider
Inbound ports	No	Usually SSH	SSH	No local nodes
Works with arbitrary local agent CLIs	Yes	Manual glue	Custom playbooks	Custom runners
Multi-machine result comparison	Built in	Manual	Manual	Pipeline-specific
Setup for an existing OneDrive user	One command	Host setup	Inventory + SSH	Provider setup
Data path	Your OneDrive	Your network	Your network	Cloud provider

CLI reference

Run

# Start watching for tasks (auto-initializes on first run)
npx -y @daomar/agentfleet run

# After first run, the shortcut is available:
agentfleet run

Options:

• --poll-interval <seconds> — polling interval, default 10
• --concurrency <number> — max concurrent agent processes, default 1
• --daemon / -d — run as a background daemon
• --log-file <path> — daemon log path, default ~/.agentfleet/agentfleet.log

Submit tasks

agentfleet submit --prompt "..." --title "..." --working-dir /path

Options:

• --prompt <text> — instruction for the coding agent, required
• --title <text> — short task title
• --working-dir <path> — working directory, default current directory
• --agent <command> — override the agent command template

Status and stop

agentfleet status              # overview of all tasks
agentfleet status <task-id>    # details for a specific task
agentfleet stop                # stop the running instance

Daemon and auto-start

agentfleet run --daemon
agentfleet install    # install auto-start on login
agentfleet uninstall  # remove auto-start
agentfleet stop

• Windows: Scheduled Task named AgentFleet, triggered on login and wake
• macOS: LaunchAgent named dev.daomar.agentfleet

Initialize (advanced)

Normally run and submit auto-initialize with OneDrive. For manual control:

agentfleet init                                    # default: OneDrive
agentfleet init --backend onedrive-business        # OneDrive for Business
agentfleet init --backend local-folder --path /shared/fleet  # local folder
agentfleet init --force                            # reinitialize

Workspace layout

~/.agentfleet/
├── config.json          # local machine config, not synced
├── processed.json       # task IDs already executed on this machine
├── agentfleet.pid       # daemon PID file
├── agentfleet.log       # daemon/auto-start log
├── tasks/ → OneDrive    # shared task JSON files
└── output/ → OneDrive   # shared per-task outputs
    └── task-001/
        ├── DESKTOP-A-result.json
        ├── DESKTOP-A-stdout.log
        └── LAPTOP-B-result.json

Prerequisites: Node.js 18+, OneDrive sync, Windows with PowerShell or macOS, and at least one coding-agent CLI or compatible local command.

Web dashboard

The hosted Pages app at agentfleet.daomar.dev/web/ is a PWA for:

• submitting tasks from desktop or mobile;
• monitoring enrolled nodes and last activity;
• browsing task history and per-machine results;
• using Microsoft Entra ID auth with PKCE and direct Microsoft Graph calls.

The page supports English and Simplified Chinese and switches automatically based on browser language.

Security model

• No inbound connectivity: no tunnels, no listening services, no exposed SSH ports.
• No AgentFleet backend: the dashboard talks directly to Microsoft Graph.
• No extra data plane: coordination stays inside the user's OneDrive/Microsoft 365 environment.
• No shared secrets in the repo: browser auth uses delegated Microsoft permissions.
• Local execution remains explicit: each enrolled machine runs its own local agent command.

Support the project

AgentFleet is free and open source. If it saves you time or inspires your distributed-agent setup:

• ⭐ Star the repository so more developers can discover it.
• 🍴 Fork it and adapt it to your own agent stack.
• 💬 Open issues/discussions with real multi-machine workflows.
• ☕ Consider sponsoring development via PayPal / WeChat Pay.

PayPal (international)	WeChat Pay (国内用户)

Supporters

hjunxu

Links

• GitHub — source, issues, discussions
• npm — package install
• Web dashboard — submit and monitor tasks
• About — architecture overview
• Contributing — development setup
• Changelog — release notes

License

MIT