OpenDAN Personal AI OS: Running AI Agents Locally

What OpenDAN Sets Out to Build

OpenDAN aims to create a personal AI operating system in the same sense that a smartphone OS organises applications: it consolidates language models, individual AI agents, and connections to personal data into one running system that lives on the user's own hardware. Rather than visiting a cloud chat interface, users have named agents that can access their local file system, email, schedule, and IoT devices once authorised.

The project defines four core abstractions. An AI Agent is driven by a language model, has its own memory, and completes tasks through natural language interaction. An AI Workflow groups multiple agents into a collaborative team to handle multi-step problems. An AI Environment is the permission layer that lets agents reach files, devices, network services, and smart contracts. The AI Marketplace is described as a one-click install path for additional agents, workflows, and models, though the README notes it was delayed from version 0.5.1 to the planned 0.5.2 release.

The practical audience for version 0.5.1 is an engineer who wants to run personal AI agents entirely on local hardware, on a PC, Mac, Raspberry Pi, or NAS, without sending conversations to a third-party service. The Docker packaging keeps the installation consistent across those platforms.

Built-in Agents and Workflows in Version 0.5.1

Four named agents ship with the 0.5.1 release. Jarvis is a personal assistant aimed at schedule and communication management, and the README positions it as a self-hosted alternative to hosted AI chat services. Mia manages personal data and sorts content into a local knowledge base; the knowledge base supports text files and common image formats, while support for other document formats is marked as incomplete in the README. Tracy serves as a private English teacher agent. The fourth agent, ai_bash, targets developers: it interprets plain-English file operation requests and translates them into shell operations, removing the need to memorise specific flags.

Connectivity to agents arrives through Telegram and email, so interactions are not limited to the local terminal. The workflow engine lets groups of agents tackle more complex tasks together; a built-in story_maker workflow combines AI agents with image generation tools to produce illustrated audio fairy tales. The model backing any agent can be switched, and LLaMA is supported as a locally run open-source alternative to hosted API models.

Installing OpenDAN via Docker

OpenDAN installs through Docker. The README lists two prerequisites: Docker at version 20.0 or later, and an OpenAI API key for the hosted model path. A local LLaMA setup removes the API key requirement, but the README does not describe that configuration in detail.

To pull the container image:

docker pull paios/aios:latest

The first launch requires interactive input for initialisation, so the -it flag is required. The README recommends mounting a local directory so personal data, chat history, and configuration persist outside the container:

docker run -v /your/local/myai/:/root/myai --name aios -it paios/aios:latest

After the named container exists, subsequent starts skip initialisation:

docker start -ai aios

For headless service mode without an interactive terminal:

docker start aios

After configuration completes, OpenDAN opens an AIOS Shell, which the README describes as similar to Linux Bash. The interface shows the active user, the current agent or workflow, and the topic. Sending a message to Jarvis and receiving a reply is the README's confirmation that the system is running. The Dockerfile in the repository uses a Python 3.11 base image, installs build tools and the MySQL client library via apt, copies source files from ./src into /opt/aios, and starts the shell entry point.

Limitations and Incomplete Features

Several features in the 0.5.1 release are incomplete or deferred. The OpenDAN Store, intended for one-click agent and model installation, was explicitly postponed to version 0.5.2. The knowledge base does not yet support most document formats beyond plain text and common image types; the README marks these as open items. The 0.5.1 system runs in an all-in-one mode, meaning all components run in a single container on a single machine; the distributed computing path, based on planned integration with the CYFS Owner Online Device OS, is described as future work.

The README itself notes the project is in very early stages with significant changes expected. There is no documentation covering rollback if an agent corrupts personal data, and the agent marketplace has not shipped. For daily personal use by non-engineers, the current state requires a significant tolerance for rough edges. For engineers building personal tooling on top of the framework, the architecture is functional but the missing marketplace and document format support are real gaps.

OpenDAN Compared to AutoGPT

AutoGPT is an open-source Python project that also runs autonomous AI agents locally or in the cloud, using a plugin system and external memory to take multi-step actions without human intervention on each step. The architectural difference between the two projects is scope. AutoGPT focuses on task automation through its plugin ecosystem. OpenDAN defines a broader personal OS abstraction, with named built-in agents that have distinct roles (scheduler, knowledge manager, language teacher), an environment layer for IoT and service access, and a workflow engine for agent teams.

AutoGPT does not ship agents tuned to personal schedule management or English tutoring. OpenDAN's Docker packaging and named container setup are more explicit about data locality than AutoGPT's default configuration. On the other hand, AutoGPT's plugin ecosystem is more developed for task automation use cases. Engineers who only need an agent that runs automated tasks should evaluate AutoGPT first. Engineers who want a named local assistant tied to their calendar and email will find OpenDAN's design more directly aligned with that goal.

Maintenance State and Licence

The repository is not archived. The last push to the main branch was on 2026-03-28. The most recent GitHub release is version 0.5.1, tagged in April 2024. The README describes the 0.5.2 roadmap as including formal OS kernel work derived from the CYFS OOD OS and the OpenDAN Store, but no 0.5.2 release had been tagged as of the last push date. Developers who want to track progress are pointed to a GitHub issue that lists system development updates.

The MIT licence places no restrictions on commercial use, modification, or distribution. The Dockerfile and full source code are in the repository, so building a custom image is possible without waiting for official releases. Anyone deploying OpenDAN to manage real personal data should treat the current state as experimental and plan for breaking changes between the current release and any future 0.5.2 tag.

Editorial conclusion

OpenDAN suits engineers who want to experiment with a local personal AI OS and named agent workflows, keeping all data on their own hardware rather than a cloud service. Users who need production reliability, a fully featured agent store, or frequent updates should check the repository carefully: the last GitHub release is version 0.5.1 from April 2024 and the last push to main was on 2026-03-28. The 0.5.2 milestone, which includes the OpenDAN Store and formal OS kernel work, had not shipped as of that date. Read the open issues before committing to it for anything beyond personal experimentation.

Frequently asked questions

Official sources

1. fiatrete/OpenDAN-Personal-AI-OS on GitHub
2. License: MIT
3. Project website
4. README
5. Releases