Self-hosted AI agent platforms: which one to run and what server it needs

Three open-source platforms cover most self-hosted AI agent building in late 2026: Dify, Langflow and n8n. Each runs on an ordinary virtual server, without a GPU, when the model sits behind an API.

Dify asks for at least 2 CPU cores and 4 GiB of memory, by its README. Flowise, once a fourth choice, reached end of life on 31 August 2026, and DataStax closed its hosted Langflow on 9 April 2026.

The facts below come from each project's own README, licence and release pages and from Hostinger's template pages, all read on 27 October 2026.

Which platform for which job

• Dify: chat apps, knowledge bases that answer from your documents, and agents, built in a visual editor. It brings its own database, vector store and code sandbox.
• Langflow: visual flows built from Python components. Its README says each flow can run as an API or an MCP server, so other tools can call it.
• n8n: workflow automation with AI agent steps and hundreds of app integrations. Our n8n hosting guide covers it in depth.

Dify

Dify runs as a group of containers: its API and workers, the web front end, PostgreSQL and Redis. By default it adds a code sandbox, a plugin service, nginx and the Weaviate vector store, so it asks for more than the others.

Its README installs it with Docker Compose 2.24.0 or later:

git clone https://github.com/langgenius/dify.git
cd dify/docker
cp .env.example .env
docker compose up -d
# then open http://your-server/install to create the admin account

The default settings serve plain HTTP on port 80. Turn on HTTPS in the .env file, or put a reverse proxy in front, before you invite anyone.

Dify's licence allows commercial use, with two limits. You may not run it as a multi-tenant service without its maker's written permission. You must also keep the Dify logo and copyright in its web interface.

Langflow

Langflow installs as a Python package. Its README asks for Python 3.10 to 3.14 and uv, and for an administrator password set before the first start:

uv venv && source .venv/bin/activate
uv pip install langflow -U
export LANGFLOW_SUPERUSER_PASSWORD='choose-a-long-password'
uv run langflow run      # starts at http://127.0.0.1:7860; sign in as langflow

Its Docker image also needs that password, by the README. Publish the port as 127.0.0.1:7860:7860 and put HTTPS in front; Docker's documentation warns that published ports are open to the world by default.

If you used DataStax's hosted Langflow, note that DataStax removed it from Astra on 9 April 2026. It points users to the open-source Langflow, the version above.

n8n as an agent platform

n8n's AI agent steps sit inside ordinary workflows, next to its app integrations. It is the lightest of the three: n8n's own example figures for embedded installs run from 320 MB to 2 GB of memory.

Our guides cover how much RAM and CPU n8n needs and self-hosting n8n with Docker, step by step.

n8n can also serve its workflows to AI assistants as an MCP server. Our guide to remote MCP servers covers that and how to lock it down.

If you used Flowise

Flowise's makers announced on 29 July 2026 that they are winding the project down. Their sunset page sets out the steps:

The code stays on GitHub under the Apache 2.0 licence, so existing installs keep working. Hostinger still offers a Flowise template, but no further updates will come from its makers.

Langflow is the closest match for Flowise's visual flows; Dify suits chat apps and knowledge bases. Move before a security problem appears that nobody will fix.

Chat apps are a different job

Open WebUI and AnythingLLM give a team a ChatGPT-style chat window rather than an agent builder. Our guide to a private chat assistant for your team covers them.

What server each one needs

• n8n: a small server is enough; n8n's own example figures run from 320 MB to 2 GB of memory.
• Langflow: its README states no minimum; give it at least 4 GB, since flows load Python libraries (our reading).
• Dify: at least 2 CPU cores and 4 GiB, by its README; 8 GB leaves room for its vector store and documents (our reading).

Hostinger's smallest plan in our records has one CPU core, below Dify's minimum. The next one up has two cores and 8 GB, by Hostinger's template pages. All three platforms leave the model to an API unless you add a GPU.

One-click options on the hosts we cover

• Hostinger offers Dify, Langflow, n8n and Flowise as one-click templates on its VPS plans. Its Dify template includes the Weaviate vector store; its Langflow template adds PostgreSQL.
• DigitalOcean's Marketplace has a one-click n8n Droplet, made by a third party, automaze.me.

A template saves the install, but you still own the updates and backups. Our AI agent hosting guide compares managed and self-managed routes.

How we checked

We read each project's README, licence and latest release on GitHub, Flowise's sunset page, DataStax's changelog and the hosts' template pages on 27 October 2026. Sizes marked as our reading are estimates, not measurements.

Our own measurements will be added here, with their dates, when our test bench runs them.

Back to AI agent hosting