AI automation tools that run natively on Windows without Docker, sorted by install route

Some AI automation tools run natively on Windows without Docker, and others expect a container or a server before they start. A quick way to tell them apart is to read each vendor's own installation page. As of October 2026, those pages sort the field into four groups: hosted plans that the vendor runs for you, self-hosted tools whose documented route is Docker Compose, tools with a second route such as npm or a Python package, and programs with an ordinary Windows installer. Langflow Desktop, AnythingLLM Desktop, Power Automate for desktop and Barside are in that last group.

Everything about other products on this page comes from their own installation documentation, as we read it in October 2026. We installed and ran none of them for this guide, so treat each line as the vendor's statement and open the linked page before you rely on it. We make Barside, a free desktop app for Windows, Linux and Mac, and it gets the same treatment here, limits included.

Which AI automation tools run natively on Windows without Docker?

Going by their documentation, four of the tools below install on Windows the way any desktop program does, and two more document a route that leaves Docker out. The table gives the route each vendor's page describes, whether Docker is part of it, and one detail from the page that is easy to miss.

Four ways an automation tool reaches a Windows computer

Hosted plans: nothing to install

A hosted plan runs on the vendor's servers, so the question of Docker never comes up. Activepieces' install options list Activepieces Cloud with the words “Fastest. Start free. We host and manage it for you.” n8n's hosting page lists n8n Cloud among its related resources, and Flowise's Get Started page points to Flowise Cloud as its hosted option.

Hosted plans are a simple way to start, and they suit jobs that must run at any hour. Plans can differ in what they count and bill. Our guide to Zapier alternatives works through that for Zapier, Make and n8n, with example volumes.

Self-hosted tools whose documented route is Docker

Deploy Dify with Docker Compose is the self-hosting quick start we opened for Dify. Its requirements list “Windows with WSL 2 enabled” next to Docker Desktop, and ask for at least 2 CPU cores and 4 GiB of RAM. That page documents Docker Compose and no other route that we found. Dify's Deploy Dify overview says “Dify runs anywhere Docker runs: a laptop, an on-premises server, or a cloud VM.” It also lists platform guides for AWS and BT Panel, and a Local Source Code Start page for people who modify or debug the platform. The source code page lists “Windows with WSL 2 enabled” with Docker Desktop too, and it starts the databases, sandbox and plugin daemon with Docker Compose.

Activepieces' install page marks one option as recommended: “Docker Compose with PostgreSQL and Redis.” It opens with a “Quick install” shell command and then says to “follow Self host (Docker) for the next steps”. It also lists a Helm chart for Kubernetes, “Only if you already run a cluster”, and templates for several cloud providers. We found no Windows-specific route and no npm route on that page.

Neither finding is a verdict on the tool. If your team already runs containers, these are the routes the vendors document first, and a server suits software that has to answer other apps at any hour.

Tools that document a second route

n8n's hosting page lists npm next to Docker Compose and says it “Installs n8n directly using Node Package Manager.” The Install with npm page calls npm “a quick way to get started with n8n on your local machine” and asks for a Node.js version between 20.19 and 24.x. It gives npx n8n to try n8n without installing it, after which you open http://localhost:5678, or npm install n8n -g and then n8n to install and start it. The page has its own Windows troubleshooting section, which sends you to Microsoft's guide for installing Node.js.

Two cautions sit on the same pages. Both say npm installs are deprecated from n8n 3.0, and the hosting page suggests Docker Compose or its one-line setup instead, so check which version you are about to install. That page also says its one-line setup “Requires Docker on Linux or macOS (or WSL on Windows).” The hosting page adds that npm “requires managing Node.js versions and dependencies yourself.” The npm page also describes a tunnel that brings requests from the web to a local instance: a callout says Docker must be installed for the tunnel even when n8n runs through npm, and the page marks the tunnel as a tool for local development and testing.

Flowise's Get Started page gives an npm quick start as well. With Node.js v18.15.0, or v20 and above, npm install -g flowise installs it and npx flowise start runs it at http://localhost:3000. A separate section covers two ways to deploy Flowise with Docker. The page does not name Windows, so it documents a Node.js route without saying how it behaves on each system.

Programs that install like any Windows app

Langflow's install page describes Langflow Desktop as “a desktop version of Langflow that simplifies dependency management and upgrades.” On Windows you double-click the downloaded .msi file and follow the install wizard. The page warns that Windows installations “require a C++ compiler that may not be present on your system”, and it names Microsoft C++ Build Tools or Visual Studio as the fix. The same page gives a Python route for versions 3.10 to 3.14 (in a virtual environment, uv pip install langflow, then uv run langflow run) and a Docker route.

AnythingLLM's Windows Installation page describes an .exe installer that you double-click, in builds for “Windows 10+ (Home, Professional - x86 64-bit)” and for ARM 64-bit. It says the desktop app “includes a built-in local LLM powered via Ollama” and that the computer needs outbound network access during installation. It also warns that “Other versions of windows (Enterprise or Server) may not work.” The Agent Flows page calls agent flows “a no-code way to build agentic skills”, with a flow editor in both the Docker and Desktop versions.

Microsoft's install page for Power Automate says to “Download and install Power Automate using an MSI installer or from Microsoft Store.” The MSI installer needs admin rights, while the Store version needs none and updates automatically. The page describes Power Automate for desktop as “the app you use to build, edit, and run desktop flows.” It says nothing about AI features, so read Microsoft's other pages before you count it as an AI tool.

Barside belongs in this group too. It is a desktop app you install on your computer, built with Electron, and it runs on Windows 10 and 11 (x64). Its agents run inside the app, and the one thing you add is an AI account in Settings → AI. The rest of this page covers what those agents do and where they stop.

Can AI agents run locally on Windows without Docker?

AI agents running locally on Windows without Docker come in two kinds, and the difference is where the AI model runs. In the first kind, the agent software and the model are both on your computer. In the second, the agent runs on your computer and sends its requests to an AI company.

AnythingLLM Desktop documents the first kind: its Windows page says the app includes a built-in local model powered via Ollama. We could not confirm on the pages we opened whether it works with no connection once it is installed. Langflow Desktop, and n8n or Flowise started through npm, also run on your own computer by their documentation. Which model each one calls is a setting to check in its docs.

Barside is the second kind. Its agents, your notes, the approval inbox and the run history are on your computer, but Barside brings no AI model of its own. Agents use the AI you connect: Claude, ChatGPT or Gemini with an API key, or OpenRouter with a one-click sign-in. A Claude or ChatGPT plan can also run agents on your computer through Claude Code or Codex, which Barside calls My plan.

Each run sends the agent's instructions and the notes it works with to that company, so it needs an internet connection. If the model has to stay on your own hardware, a tool with a local model fits better than Barside.

What Barside's agents and saved workflows do as a desktop app

Barside keeps its agents under Agents & approvals, which has six tabs: Inbox, History, Scheduled, Meeting, Research and Workflows. Each agent has its own way to start, and none of them starts from another web app.

Proposals collect in Agents & approvals → Inbox. There you can press Edit to change a proposal, then Approve or Refuse, and the History tab records each decision. Barside never sends an email itself: approving one opens your email app with a draft, and you press send there. The AI agents guide covers each agent in detail.

On Windows, a saved workflow can also work in other apps. The assistant reads an app's accessible controls through Windows UI Automation and can focus, click, type or press keys there. Each of those steps first shows a Review computer action card with Approve this action and Cancel action. Apps with custom-drawn screens, and apps running as administrator, may not work.

How to set up a first agent on Windows

1. Download Barside and install it on Windows 10 or 11.
2. Open Settings → AI and connect Claude, ChatGPT, Gemini or OpenRouter. Without an AI, agents cannot run.
3. Open Agents & approvals in the sidebar and choose the Scheduled tab.
4. Choose the Morning briefing preset, which runs on weekdays at 8:30 AM. Change the Instructions if you like. An example: “Put anything due this week first.”
5. Keep A new note each run under Results, pick a Workspace, leave Run on schedule ticked and press Create agent.
6. Press Run now to see the first note, then open the Inbox for anything the agent proposes.

To repeat a job of your own, give the assistant a request once and let it finish. Under its last reply, choose Save as workflow and then Save workflow. The saved workflows guide shows how to run it by name with Alt+Shift+Space, and how to give it a schedule that asks you before each run.

A scheduled agent needs Barside to be running at its time. Closing the main window leaves Barside in the tray, and Start with Windows, under Windows & docking in Settings, opens it quietly when you sign in. That setting is off by default.

Jobs that fit a desktop agent

Every person, request and question below is an example, written to show the shape of the job.

• Mara, a freelance bookkeeper, keeps the Weekly review preset, which runs on Fridays at 4:00 PM. It writes a note from her open tasks and reminders, and each reminder it suggests waits in the Inbox.
• Tomas, an account manager, ticks After each call transcript, prepare a follow-up on the Meeting tab. After a call, a Follow-up checklist and a recap email wait for him. He presses Edit, adds the client in To, approves, and sends the draft from his own email app.
• Lin, a graduate student, types an example question on the Research tab: “How do heat pumps perform in cold climates?” She gets a note with numbered sources.
• Dev, a recruiter, saves an example request as a workflow: “Collect every tab of this browser window into a new note called Friday research.” The next week he runs it by name from the search bar.

Where a server tool is still the better choice

A desktop app does not suit every kind of automation, and it helps to know that before you install anything.

• Jobs that start in another web app. Barside has no app-to-app connectors or webhooks, so nothing fires on a new form entry or a new email. A hosted plan or a self-hosted server fits those jobs. n8n's npm page presents a local install as a way to get started, and it says the tunnel that brings outside requests to a local instance is unsafe for production.
• Jobs that must run at any hour. Barside's scheduled runs never start while the computer is off, asleep or locked. Afterwards, only the latest missed run from the last 12 hours is made up. A server that is always on has no such gap.
• Runs that should finish with nobody watching. In Barside, each scheduled workflow run waits for Approve, and each action in another app asks again. That suits work you want to check, and it slows work you want hands-off.
• A model on your own hardware. Barside brings no AI model. AnythingLLM Desktop's Windows page documents a built-in local one.

Barside Plus narrows the second gap. For US$20 a month or US$200 a year plus sales tax, cloud agents run your scheduled agents on Barside's server while your computers are off, with your own AI key. Saved workflows and the meeting agent stay on your computer.

A mix is a sensible setup: one always-on tool for app-to-app jobs, and a desktop app for the briefings, follow-ups and research you want to approve. The comparison of human-in-the-loop automation tools looks at how approval works in each.

Limits and details

• Your AI account pays. Barside is free, and the AI company you connect bills what the agents use. Settings → AI has an AI spending limit for what each computer uses.
• Numbers. You can keep up to 50 scheduled agents and 100 saved workflows, with 40 recorded steps in a workflow. One AI-backed agent runs at a time while the others wait.
• Schedules. Times are your computer's local times, and Barside checks the schedule every 30 seconds.
• Other apps differ by system. On Windows the assistant reads and presses controls through Windows UI Automation. On Linux it can focus and minimize windows and press keys but cannot read the controls inside other apps, and on Mac it uses System Events once you allow Barside under Accessibility.
• Inside Barside, nothing differs. Agents that work with notes, tasks and Reading behave the same on Windows, Linux and Mac.

Privacy: what stays on your computer

Your scheduled agents and saved workflows are stored in your library on your computer, and the Inbox and its History stay on each computer. When an agent runs, its instructions and the notes, transcripts or pages it works with go to the AI company you connected, whose data policy applies. Barside sends nothing to its own servers to run a saved workflow.

Notes and workspaces you untick in Library & briefing → Sources are never given to an agent. The research agent sends your search to DuckDuckGo and opens the pages it reads without your browser's cookies or sign-ins. Cloud agents are different: they read your synced notes on Barside's server, and that server can read synced notes. Read how agents use your data and the guide to local-first storage and backups.

Works well with

If you would rather keep working in Claude Desktop, ChatGPT or Cursor, AI connections give those apps tools for your notes and tasks. For the app itself, see how the native desktop app docks to a screen edge.