Skip to content
Local-first desktop agent

Your AI workforce.
On your machine.

An autonomous agent beside a real, multi-tab browser it can read and drive — and, when you allow it, your whole computer. Point it at any model from any provider with your own key. Everything stays local. Open source.

Windows · macOS · Linux — bring your own API key

AI WorkForce OSbuild

Summarize this page.

Ask AI WorkForce OS to do something…
localhost
localhost
1
2
Export3

Try a task

  • Local-first
  • Bring your own key
  • No accounts
  • No telemetry
  • MIT open source
  • Windows
  • macOS
  • Linux
The third path

Between a chat box and someone else's cloud.

Assistants describe steps but can't touch your screen. Cloud agents take your data, rent you one model, and run the browser on their server. AI WorkForce OS is the other path.

Chat assistants

  • Answer questions, but can't act
  • Describe the steps — you do them
  • No access to your screen or files

Cloud agents

  • Run on someone else's server
  • Your data leaves your machine
  • Rented to one vendor's model

AI WorkForce OS

The third path
  • A real embedded browser it drives
  • Any model, any provider, your key
  • Everything stays on your machine
Why it's different

Built to do the work, not describe it.

Six things every other category gives up. Each one maps to a real mechanic in open, auditable source — not a tagline.

  • Local-first & private

    Runs as a normal desktop app. Keys live in a JSON file on your disk. No accounts, no relay server, no telemetry — nothing leaves your machine except the calls to the provider you chose.

  • No lock-in, any provider

    Claude, OpenAI, Gemini, OpenRouter, any OpenAI-compatible endpoint, or local Ollama — switchable per agent. One runtime dependency. Nothing proprietary to be trapped by.

  • It acts, not just chats

    A 21-tool browser surface — read, click, fill, work native and custom dropdowns, scroll, navigate, manage tabs, screenshot, upload. On desktop it adds 15 tools to drive your whole computer.

  • Automation that sticks

    Capture a routine once and replay it: quick linear step-lists, or full visual node graphs run by a deterministic engine — not the model improvising each run. Graphs can be scheduled.

  • MCP, both ways

    Pull external tools in as an MCP client, and expose its own browser out as an MCP server — so hosts like Claude Code can drive its tabs. A tool consumer and a tool provider.

  • Open source, MIT

    Read every line, fork it, ship it. The agent brain is plain JavaScript you can audit — transparency is the whole pitch. MIT © Rises Solutions.

How it works

Three steps. No setup theatre.

Connect a model, ask in plain language, and watch it act — every action an inspectable chip.

  1. 01

    Connect a model

    Paste an API key for Claude, OpenAI, or Gemini — or one-click sign in with OpenRouter for free models, no card. Point it at any OpenAI-compatible endpoint, even local Ollama.

    • Anthropic
    • OpenAI
    • Gemini
    • OpenRouter
    • Ollama
  2. 02

    Ask in plain language

    Type what you want done in the composer — "export this month's CSV", "summarize this page", "fill this form from my notes". Switch personas with one keystroke.

    • build
    • plan
    • explore
    • reason
  3. 03

    It does the work

    The agent reads the embedded tab and takes the next concrete step, showing each action as an inspectable chip you can expand or stop. Save it as a workflow to run again — or on a schedule.

    • Read page ✓
    • Click ✓
    • Type ✓
    • Done ✓
Under the hood

Concrete tools, not vague promises.

Every capability maps to a real mechanic in the source. That precision is the pitch.

A full browser surface the agent can drive

Read the active tab as text plus an indexed map of interactive elements, then act: click, fill, select native and custom/ARIA dropdowns, scroll, press keys, find text, navigate, go back, and open/switch/close tabs. It can screenshot — plain or with numbered element boxes for vision models — click by coordinate, and upload files.

read_pageclickfillselect_optionscrollpress_keysfind_textnavigatego_backopen_tabswitch_tabclose_tabscreenshotset_of_marksclick_xyupload_file
Bring your own key

Any model. Any provider. Your key.

Claude, OpenAI, Gemini, OpenRouter, any OpenAI-compatible endpoint, or local Ollama — switchable per agent.

  • ClaudeAnthropic
  • OpenAI
  • GeminiGoogle
  • OpenRouterone-click sign-in
  • Ollamalocal
  • OpenAI-compatibleany endpoint

Switch model per agent. Your key, your endpoint.

Security & privacy

Transparency is the product.

Keys on disk, no telemetry or relay, a permission gate on every powerful action, page text treated as untrusted, and MIT-auditable source. The whole pitch is that you can verify it.

  • Keys stay on disk

    API keys live in ai-workforce-os-store.json in your OS user-data dir and only ever go to the provider you picked. No accounts, no relay, no telemetry middleman.

  • Ask / Auto permission model

    Browser actions can be set to Ask, with per-site Blocked rules until you trust a site. An explicit site choice beats every other rule.

  • System control is gated

    Driving your whole computer defaults to Ask — every shell, file, and input action waits for your in-chat approval. Set it to Off to keep the agent inside the browser.

  • Page text is untrusted

    The agent treats page content as untrusted and is told never to enter credentials or payment details — so you stay in the loop on logged-in sites.

  • MIT-auditable

    The agent brain is plain JavaScript under an MIT license. Read every line, fork it, and verify exactly what it does — nothing is hidden in a binary.

  • Runs in your process

    No code executes on a vendor's server. The capability is local to your machine; only the prompts and results you send flow through the model you chose.

How it compares

Three categories, compared honestly.

Not a competitor takedown — a comparison of approaches across what actually matters for a local agent.

Capability comparison across AI WorkForce OS, Cloud agents, Chat assistants.
CapabilityAI WorkForce OSCloud agentsChat assistants
Runs on your machineYesNoNo
Your own keys, any modelYesNoNo
Acts on the screenYesYesNo
Drives your whole computerYesNoNo
Open source / auditableYesNoNo
No data sent to a vendor's serverYesNoNo
Pricing

Pricing that scales with use.

Paid plans route model calls through the managed gateway so usage can be metered and billed. Free, plus your own key, stays fully local.

Free / Open-Source

The full desktop app, forever. Bring your own key.

$0/mo

Free forever, no account needed

  • The complete AI WorkForce OS desktop app (MIT)
  • Bring your own key — any provider, fully local
  • 21 browser tools + 15 desktop tools + workflows
  • No account required, no telemetry
Most popular

Plus

Included usage through the managed gateway.

$19.00/mo

5,000 credits/mo

  • Everything in Free
  • Managed metered gateway — no provider keys to manage
  • Plus-tier models (Claude Sonnet/Haiku, GPT-4o, Gemini Flash)
  • Usage dashboard + monthly credit grant
  • Email support
Subscribe

Continue to checkout after sign-in.

Pro

More credits and the top-tier models.

$49.00/mo

25,000 credits/mo

  • Everything in Plus
  • Pro-tier models (Claude Opus, OpenRouter Auto)
  • 5× the monthly credits
  • Priority support
  • Team foundations (coming soon)
Subscribe

Continue to checkout after sign-in.

Prices and monthly credit amounts shown are configurable placeholders — set your real economics (and matching Stripe prices) before launch.

Get started

Download and point it at your model.

Free OpenRouter sign-in or any provider key. No account, nothing to configure on a server.

Windows

.exe

NSIS installer

Download Windows

macOS

.dmg

Universal (Intel + Apple Silicon)

Download macOS

Linux

.AppImage

Portable, no install

Download Linux

Bring your own API key — free OpenRouter sign-in or any provider key. Node 18+ to build from source.