Open-source AI agents: deployment, models and costs

Choose by the workflow you can operate, not just by whether the repository is public. These four existing directory products have MIT or Apache-2.0 code licenses; they are different kinds of tools, not interchangeable agents.

Official sources checked November 5, 2026. This is a documentation-based selection guide, not a hands-on benchmark or a ranking of output quality.

Four starting points by workflow

Aider

Terminal coding with Git diffs

Install the Python CLI and connect a supported cloud or local model. Review patches and run your own tests.

Code license: Apache-2.0

Official repository →

Langflow

Visual flows and APIs

Build flows visually, extend them with Python components and expose APIs or MCP tools. Self-hosting requires maintaining the server and dependencies.

Code license: MIT

Official repository →

smolagents

Custom Python agents

Use CodeAgent for Python actions with hosted or local models. Choose a sandbox and supply your own tools and evaluation tasks.

Code license: Apache-2.0

Official repository →

OpenHands

Self-hosted coding workflows

Run coding workflows with local or remote execution backends. Check the terms of each connected agent and any optional cloud service separately.

Code license: MIT

Official repository →

Hosted API or local model?

A hosted API reduces local model setup, but requests may leave your machine and incur usage charges. A local model changes the hardware and maintenance requirements. Check the actual model, context length and concurrency before choosing RAM or GPU capacity; there is no single hardware minimum for this whole list.

Budget for model inference, execution or hosting, storage and maintenance time. A free client or permissive code license does not make those costs disappear. Self-hosted software can still call external APIs.

Public source code is not always a permissive license

n8n uses the Sustainable Use License for its non-enterprise code, with internal-use and redistribution conditions. Dify uses an Apache-based license with additional conditions for multi-tenant services and frontend branding. Review the actual license for your deployment before treating either as an unrestricted open-source substitute.

n8n · n8n license · Dify · Dify license

Run one controlled pilot before committing

  1. Choose one representative task and preserve its input and expected result.
  2. Start with limited permissions; inspect the model endpoint and sandbox before providing private data.
  3. Record model charges, runtime, manual corrections and hosting overhead for the same task.
  4. Review generated changes and require human approval before external writes or production actions.

Compare coding agents → · Explore developer infrastructure →