Open-source automation and AI assistants: n8n, Dify, Browser Use, MiroFish

What this guide covers

• What each of the four tools does, in plain words, according to its own README and documentation
• One small first project for each tool that a small business can try
• What you need to run it: the vendor's cloud or your own server, API keys, skills
• What each licence file allows and where it draws a line
• A comparison table and a short answer to which tool to start with
• Safety habits for API keys, personal data and automated actions

Four tools at a glance: public code, four different licences

All four publish their code, but the licences differ. Browser Use is under the MIT License and MiroFish under the GNU AGPL version 3, both standard open-source licences. Dify uses a modified Apache License 2.0 with extra conditions. n8n has its own Sustainable Use License: its documentation says the company does not call n8n open source, because the licence limits commercial use, and prefers the term fair-code.

n8n: workflows that connect your apps

The README calls n8n a fair-code platform for building AI agents and workflows: you draw a process on a visual canvas and add code where ready blocks fall short. According to the documentation, a workflow can start from a webhook or on a schedule, branch on conditions, run your own JavaScript or Python, and pause until a person approves an action an AI agent wants to take.

A first project: a website enquiry arrives by webhook, n8n adds it as a row to Google Sheets, a language model drafts a reply, and a task with a reminder appears in your task app. A person reads the draft and sends it.

What you need: n8n Cloud, a paid subscription with a 14-day free trial, or the free self-hosted Community edition. According to the documentation, self-hosting requires technical expertise, and the server and its upkeep are your responsibility. AI steps need a model provider's API key.

The licence: LICENSE.md allows using or modifying the software only for your own internal business purposes or for non-commercial or personal use, and providing it to others only free of charge for non-commercial purposes. Files marked .ee fall under the separate n8n Enterprise License. According to n8n's licence FAQ, running your business on it and charging clients for building workflows is allowed; hosting it as a service where clients build their own workflows is not.

A poor fit if nobody can look after a server and you do not want the cloud plan, or if you plan to resell n8n itself: for that the documentation requires a separate commercial agreement.

Dify: an AI app that answers from your documents

The README describes Dify as an open-source platform for developing LLM apps: a visual canvas for AI workflows, text extraction from PDFs and other common document formats, agents, logs and an API for every app. In the documentation your data lives in knowledge bases, and a finished app can be shared as a web link, embedded in a website as a chat widget or called over the API.

A first project: an internal helper for staff. Load your price list, delivery terms and most frequent customer answers into a knowledge base, build a chat app on it and share the link with the team. Run your twenty most common questions through the retrieval test described in the documentation before anyone relies on the answers.

What you need: Dify Cloud, or the self-hosted Community Edition started with Docker Compose; the README gives a minimum of 2 CPU cores and 4 GiB of RAM. Either way you connect a model provider, usually with an API key and usage billed by that provider.

The licence: the LICENSE file says Dify is under a modified Apache License 2.0 and may be used commercially, including as a backend for other applications. A commercial licence from the producer is required in two cases: operating a multi-tenant environment from the Dify source code without written authorisation (one tenant is one workspace), and removing or changing the logo or copyright information in the Dify console or applications when you use its frontend.

A poor fit for a service where each client gets a separate workspace, unless you obtain the commercial licence. The documentation says a knowledge base makes answers less prone to hallucinations; that is not the same as free of them, so a person still checks prices and deadlines.

Browser Use: an agent that works in a browser

Browser Use is an AI browser agent: you describe a task in a sentence, and the agent opens a browser, works through the pages and returns an answer. The README offers three routes: a hosted cloud, a command-line tool that gives browser access to an agent you already use, and an open-source Python library run from your own code.

A first project, read-only on purpose: once a week the agent opens five public supplier pages you name and writes price and stock status into a table, which you then compare with the pages. Check that each site's terms of use allow automated visits.

What you need: for the library, Python 3.11 or newer and a model API key in an .env file; the browser can be local or hosted by the vendor. The README says the library is free, while model usage and hosted browsers are charged separately. Someone must be comfortable with Python and a terminal.

The licence: the LICENSE file is the MIT License. It permits using, copying, modifying, distributing and selling the software, provided the copyright notice and the permission text stay in all copies. The software comes as is, without warranty.

Be careful with logins. The README shows how to reuse your Chrome profile, and then the agent acts inside your accounts. Start with a clean profile and leave payments, messages and deletions to a person.

MiroFish: a simulation of how people might react

The README presents MiroFish as a swarm intelligence engine. You upload seed materials, for example a report, and describe in plain language what you want to foresee. The engine builds a knowledge graph, generates personas, runs a simulation in which these AI agents interact and writes a report. The published demos are a public opinion simulation and the lost ending of a classic novel.

A first project: before announcing a price change, upload the draft announcement and a description of your customer groups, without names or contacts, and ask how the discussion may unfold. Put the objections you had not thought of to five or ten real customers.

What you need: Node.js 18 or newer, Python 3.11 or 3.12 and the uv package manager, or Docker. Two keys are required: a language model API in the OpenAI format and Zep Cloud. The README warns of high model consumption and suggests starting with fewer than 40 simulation rounds. This is a developer's tool.

The licence: the LICENSE file is the GNU Affero General Public License, version 3. You may use, change and share the program. A modified version you pass on must stay under the same licence, and section 13 says that if you modify the program and let people use it over a network, you must offer them your version's source code at no charge.

The main caution: the README speaks of predicting anything, yet on 4 October 2026 it contained no accuracy figures and no comparison of past simulations with real outcomes. A simulation gives hypotheses, not proof of demand: it shows what language models imagine people would say. Treat the report as questions for real customers, never as a reason to order stock.

Which one first?

Start from the task, not the tool. Write down the process as you do it by hand today and pick the one tool that matches.

One tool, one process, two weeks. If the automation still needs daily repair after that, the process was not ready to be automated.

• A routine repeats between apps (enquiries, tables, reminders): n8n, on the cloud trial, so no server is involved.
• People keep asking questions whose answers sit in your documents: Dify.
• The work exists only as clicks on a website with no export or API, and you have a developer: Browser Use.
• You have a specific question about public reaction and a budget for model usage: MiroFish, as an experiment and last on the list.

Safety: keys, personal data and checking the machine's work

All four tools send text to a language model and act on your behalf. Four habits keep that under control.

Read the licence file again before any commercial use, because terms change.

• API keys. Keep a key only in the tool's credentials settings or an .env file, never in a prompt, a chat or a screenshot. Use a separate key per tool and set a monthly spending limit where the provider offers one, for example $20.
• Personal data. Whatever a workflow sends to a hosted model leaves your systems. Remove names, phone numbers and addresses where the task works without them, and check what your data protection law and the provider's terms require.
• Access. Give each tool the narrowest access that works: one table rather than the whole drive, a browser profile without saved logins.
• Checking. Let the automation prepare and a person approve until you have reviewed a few dozen real cases. Anything that sends, pays or deletes stays behind a human approval step.