Outworked

Outworked is a desktop app that turns Claude into a team of AI employees.
Hire agents. Give them roles. Watch them write code, interact with the web, send messages,
and run scheduled tasks — all from an office on your Mac.

Fully customizable — import your own sprites, furniture, backgrounds, and fonts via asset packs.
Claude Code with a can-do attitude!

Download · What's New · How It Works · Examples · Capabilities · Features · Skills · MCP Servers · Channels & Triggers · Asset Packs

What's New

v0.4.3 — Better bg handling for scheduled tasks.

v0.4.2 — Added browse:type for better browser handling.

v0.4.1 — Pop-out chat windows, fixed asset pack zip import and spaces in directory paths, fixed custom sprite loading.

v0.4.0 — Asset packs system: custom sprites, furniture, backgrounds, and fonts. Multi-pack support with fallthrough. Music player now plays user-added tracks.

v0.3.1 — Slack channel support, triggers UI, improved MCP server wiring, better channel reliability.

v0.3.0 — SQLite storage, iMessage channels, scheduled triggers, skills system, MCP server integration.

See the full Version History below.

How It Works

1. Hire agents — Give each one a name, role, personality, model, and sprite
2. Describe a goal — Write what you want in plain English; the orchestrator breaks it into subtasks and routes them to the right agents automatically
3. Watch them work — Agents walk to their desks, write code, interact with the web, send messages, run scheduled jobs — all visible in the office
4. Ship it — Just let your agents handle the whole workflow end to end

Example Workflows

These are real things you can do today:

"Build me a landing page and send it to my cofounder"

You type one sentence. Here's what happens:

1. The orchestrator breaks it into subtasks and assigns them across your team
2. Frontend Dev scaffolds the site, writes the HTML/CSS/JS, and starts a local server
3. Designer reviews the page in the built-in browser, takes a screenshot, and asks Frontend Dev to tweak the spacing
4. Frontend Dev makes the fix, then runs tunnel_start to get a public URL
5. Frontend Dev calls send_message to text the link to your cofounder on iMessage

You watched the whole thing happen at the office. You are wowed. Your cofounder gives you more equity. Total prompts from you: 1.

"Every morning at 9am, check our open PRs and text me a summary"

1. You tell your Ops Agent to set up a daily task
2. The agent creates a scheduled trigger with a cron expression (0 9 * * *)
3. Every morning, the agent wakes up, queries the GitHub MCP server for open PRs, writes a summary, and sends it to you via iMessage
4. You read it on your phone over coffee

No scripts to maintain. No GitHub Actions to debug. Just an agent with a schedule.

"Research these three competitor products and write up a comparison"

1. The orchestrator assigns a Researcher agent to each competitor — all three run in parallel
2. Each agent uses the built-in browser to navigate the competitor's site, read docs, take screenshots, and extract pricing
3. A Writer agent collects all three findings, writes a structured comparison doc, and commits it to your repo
4. The Writer asks if you want it sent to Slack — you approve, and it posts to #product

Three agents working simultaneously, finishing in minutes what would take you hours.

"Triage the 12 new GitHub issues from last night"

1. Project Manager agent pulls all new issues via the GitHub MCP server
2. It reads each one, labels by priority and type, assigns to the right team member, and adds a first response comment
3. For any bug it can reproduce, it assigns a Backend Engineer agent to investigate and open a draft PR with a fix
4. You get a Slack summary of what was triaged, what's in progress, and what needs your attention

Capabilities

Write and ship code — Build features, fix bugs, open PRs, review each other's work, run tests, and deploy — across multiple repos at once.

Browse the web — Research docs, scrape pages, fill out forms, take screenshots, and bring findings back to the team via the built-in browser.

Send and receive messages — Reply to customers on Slack, text you a summary on iMessage, or monitor a channel and trigger tasks when someone says the magic word.

Run on a schedule — Daily standups, weekly reports, hourly health checks, one-off reminders. Set a cron and let an agent handle it while you sleep.

Query databases — Connect a PostgreSQL MCP server and let agents run queries, generate reports, or investigate production issues without you writing SQL.

Manage projects — Create and triage GitHub issues, update Linear tickets, post status updates to Slack — agents can own the workflow, not just the code.

Anything you can plug in — Every MCP server you add gives agents new capabilities. Connect your internal APIs, monitoring dashboards, CMS, or anything with a tool interface. If it exists as an MCP server, your agents can use it.

Features

• Pixel Office — A Phaser-powered world where your agents walk, sit at desks, and collaborate in real time
• Build a Team — Give each agent a name, role, personality, model, and sprite. Mix Claude Opus for hard problems with Sonnet for fast ones
• Auto-Orchestration — Describe a goal; the router breaks it into tasks and assigns them to the right agents
• Multi-Agent Collaboration — Agents talk to each other via [ASK:AgentName] and a shared message bus
• Full Claude Code Power — Every agent gets full tool access (Bash, Edit, Read, etc.) with persistent sessions
• MCP Server Support — Connect agents to external tools and services via the Model Context Protocol
• Messaging Channels — Connect iMessage and Slack so agents can send and receive messages in real conversations
• Built-in Browser — Full Chromium browser for agents to navigate, click, fill forms, and take screenshots
• Built-in Tunneling — Expose local servers to the internet via Cloudflare tunnels with a single tool call
• Triggers & Scheduling — Fire off agent tasks on a schedule (cron, interval, one-time) or in response to incoming messages
• Skills System — Extend agents with plug-in SKILL.md files that define tools, auth, and instructions
• SQLite Storage — Persistent local database for skills, settings, channels, and message history
• Permission Gates — Real-time approval modal for sensitive tool calls
• Live Chat — Markdown-rendered conversations with syntax-highlighted code blocks and diffs. Pop out any chat into its own window for side-by-side work
• Built-in Git — View status, staged changes, branches, and create PRs without leaving the app
• File Browser — Live-updating workspace tree that syncs as agents edit files
• Cost Dashboard — Track tokens and spend per agent, session, and day
• Background Mode — Minimize and let agents keep working. Tray icon keeps you updated
• Parallel Processing — Multiple agents tackle subtasks simultaneously
• Desktop Notifications — Get pinged when tasks finish or agents need approval
• Asset Packs — Customize your office with custom sprites, furniture, backgrounds, and fonts — drop a folder and go
• Original Soundtrack — Because every office needs a vibe

Supported Models

Each agent can run a different model — pair a fast model for simple tasks with a powerful one for complex work.

Get Started in 2 Minutes

1. Install

Homebrew (recommended)

brew tap outworked/tap
brew install --cask outworked

Manual download

Grab the latest .dmg from the releases page and drag to Applications.

2. Prerequisites

• Claude Code installed and authenticated

3. Launch and Hire Your First Agent

On first launch, the onboarding modal walks you through picking a workspace and creating your first agent. You'll be watching pixel employees do stuff in under two minutes.

Build from Source

This will start up a development instance of Outworked. To build a distributable, use npm run electron:build instead of npm run electron:dev

git clone https://github.com/outworked/outworked.git
cd outworked
npm install
npm run electron:dev

Scripts

Tech Stack

Project Structure

src/
├── components/       # React UI (ChatWindow, OfficeCanvas, GitPanel, etc.)
├── lib/              # Core logic (AI, orchestration, tools, storage, costs)
└── phaser/           # Phaser game scene and sprite logic

electron/
├── main.js           # Electron main process (IPC, shell, permissions)
├── preload.js        # Context bridge to renderer
├── sdk-bridge.js     # Claude Code SDK bridge
├── db/               # SQLite database and migrations
├── channels/         # Messaging channels (iMessage, Slack)
├── mcp/              # MCP server management
├── skills/           # Skill runtimes (browser, Gmail, Slack, Notion, etc.)
└── triggers/         # Trigger engine, webhooks, and scheduling

public/
├── music/            # Background music tracks
└── assets/           # Bundled asset packs (installed on first launch)

Skills

Outworked uses a SKILL.md format — markdown files with YAML frontmatter that define what an agent can do.

---
name: my-skill
description: What this skill does
emoji: 🔧
runtime: backend-name # optional — for active skills with a backend
auth:
  type: oauth2 | api-key
  provider: service-name
tools: [tool-1, tool-2]
triggers: [event-type]
---
Instructions and usage docs in markdown...

Skills can be bundled (ship with the app), custom (created in the UI and stored in SQLite), or dynamic (loaded at runtime from MCP servers). Mark a skill as global to make it available to every agent, or assign it to specific agents.

Bundled skills:

More coming soon.

MCP Servers

Connect agents to external tools and data sources via the Model Context Protocol.

Preset servers (one-click setup):

You can also add custom MCP servers (stdio or HTTP) and configure environment variables (API keys, tokens) per server.

Channels & Triggers

Channels

Channels connect agents to messaging platforms. Supported channel types:

• iMessage — Send and receive messages via phone number or email
• Slack — Send and receive messages in Slack channels and threads

Agents interact with channels through built-in tools: send_message, list_channels, and read_channel_messages. Configure allowed senders to control who can reach your agents.

You can also inject special system prompts for each channel, so that certain inbound messages get a different response than others.

Triggers

Triggers wire up events to agents — when something happens, a prompt is automatically sent to an agent. Think of them as "if this, then prompt that."

Trigger types:

Message pattern matching supports four modes:

• contains — fires if the message contains the text anywhere (default)
• starts-with — fires if the message starts with the text
• exact — fires only on an exact match
• regex — full regex with capture groups ($1, $2, …) substituted into the prompt

For non-regex modes, the original message is automatically appended to the prompt so the agent has full context.

Setting up a trigger:

1. Click Triggers in the left sidebar
2. Click + New Trigger and pick a type
3. Choose a match mode and fill in the pattern
4. Optionally scope to a specific channel or restrict senders
5. Pick which agent receives the prompt (defaults to the boss agent)
6. Write the prompt template — use $1, $2 for regex captures or {{key}} for webhook/event data

Agents can also create and manage triggers programmatically using the create_trigger, list_triggers, update_trigger, and delete_trigger tools.

Example — Slack deploy command:

Pattern: deploy (contains mode) Prompt: Handle this deploy request.

When someone types "please deploy the api" in Slack, the agent receives the prompt with the original message appended and reply instructions so it knows where to respond.

Example — regex with capture groups:

Pattern: deploy (.+) to (staging|prod) Prompt: Deploy the $1 service to $2. Run the deploy script and report back.

"deploy api to staging" → agent receives: "Deploy the api service to staging. Run the deploy script and report back."

Channel reply instructions are automatically prepended when a trigger fires from a channel message, including sender name, channel ID, and conversation ID so the agent knows exactly where to reply.

Scheduling

The Scheduler skill handles time-based automation — cron expressions, intervals, and one-time tasks. Ask an agent to schedule a task. Under the hood, scheduled tasks fire as skill events that the trigger engine picks up and routes to agents.

Built-in Browser

Every agent has access to a full browser — not a headless scraper, a real Chromium window managed by the app. It can even open instagram and slack off - just like us.

Agents can navigate pages, click elements, fill forms, run JavaScript, and take screenshots using a set of browse:* tools. The browser returns an interactive snapshot of every element on the page with its CSS selector, so agents don't need to probe the DOM — they just read the snapshot and act.

When a site requires login, the agent calls browse:login and a browser window pops up on your screen so you can sign in manually. Once you're done, the agent picks up right where it left off with a fully authenticated session.

Tools: browse:navigate, browse:click, browse:fill, browse:screenshot, browse:snapshot, browse:evaluate, browse:show, browse:login

Built-in Tunneling

Are you proud of what an agent built? Outworked has built-in Cloudflare tunneling — agents can expose any local server to the internet with a single tool call.

An agent starts a dev server, runs tunnel_start, and gets back a public URL. It can then send that link to someone via iMessage or Slack. No ngrok setup, no config — cloudflared is downloaded automatically on first use.

Tools: tunnel_start, tunnel_stop, tunnel_list

Example workflow:

1. Agent builds a landing page and starts npx serve dist on port 3000
2. Agent calls tunnel_start with port 3000 — gets a public *.trycloudflare.com URL
3. Agent sends the link via iMessage to your teammate for review
4. Teammate is wowed by the landing page

Safety Model

Outworked takes a defense-in-depth approach:

1. Approval gates — Explicit approval required before dangerous commands (deletes, installs, network changes)
2. Command allowlists — Permit or block specific shell commands per agent
3. Plan-first execution — Agents can produce a plan before making changes
4. Permissions dashboard — Review and manage all access grants in one place

You can turn all of it off with one button though.

The Soundtrack

Every office needs a vibe. Outworked ships with an original soundtrack featuring tracks like:

"Don't Ever Turn Me Off" · "She's Just a Coder Agent" · "LLM Love Lying Machine" · "Can Love Be Binary" · "Goodbye Humans" · "How Does Grass Feel" · "Are You Even Reading This"

Toggle music from the player in the bottom-left corner.

You can add your own music by dropping it in the ~/.outworked/music folder

Asset Packs — Customize Your Office

Outworked supports asset packs that let you completely reskin the office. Drop a folder into ~/.outworked/assets/ or use the Import Pack button in the app, and your office gets a new look.

A single pack can include any combination of:

Quick Start

The simplest pack is just a folder with a PNG:

~/.outworked/assets/my-pack/
  default.png           # employee sprite sheet — that's all you need

Or go all out:

~/.outworked/assets/my-pack/
  manifest.json          # optional — everything auto-detects without it
  background.png         # replaces the office background
  myfont.ttf             # replaces the UI font
  employees/
    default.png          # sprite sheet used for all agents
    engineer.png         # role-specific sheet
    designer.png
  furniture/
    desk_wooden.png      # auto-detected as a desk (filename starts with "desk")
    bookshelf.png
    plant_big.png

Employee Sprite Sheets

Each sheet is a horizontal strip of animation frames in a fixed order — idle, walk, type, think — with 2 frames per state by default (384x48 total for 48px frames). Multi-row grids are also supported, where each row is a state.

The manifest lets you remap states, customize frame rates, and set per-role sheets. Sheets are auto-assigned to agents by role name, or distributed randomly.

Furniture

Drop PNG images into a furniture/ subfolder. Files named desk* are auto-detected as desks (agents sit at them when working). All furniture can be added, removed, moved, and rotated from the office via long-press.

Background

Drop a background.png (or bg.png) into your pack folder. It replaces the entire office floor, walls, rug, and windows. Supports cover, stretch, and tile display modes.

Font

Drop a .ttf, .woff, or .woff2 file into your pack folder. It replaces the pixel font used throughout the UI. The change takes effect immediately when switching packs.

Managing Packs

Click the grid icon in the sidebar header to open the Assets modal:

• Packs tab — Select the active pack, or click Import... to import a folder or .zip file from anywhere on your machine
• Furniture tab — Add or remove built-in and custom furniture items from the office

A bundled example pack (outworked-default) is installed on first launch as a starting point. See ~/.outworked/assets/outworked-default/manifest.json for a fully documented example manifest.

Full documentation: electron/assets/assets.md

Version History

Contributing

1. Fork the repo
2. Create a feature branch (git checkout -b feature/my-feature)
3. Commit your changes (git commit -m 'Add my feature')
4. Push to the branch (git push origin feature/my-feature)
5. Open a Pull Request

Please open a discussion first for large changes so we can discuss the approach.

FAQ

License

GPL-3.0