A Docker-based Perplexity Computer alternative can bring web search, terminal and file access, and app integrations into a self-hosted AI workspace. But running the workspace on your computer does not mean the AI model processes requests there: Jayric Maning reports using a paid inference API when local models were too slow for his serious tasks.
Check where inference happens separately from where the interface runs. A locally hosted AI workspace can still send model requests to an external API.
What does the Docker setup provide?
Maning’s setup combines Docker Desktop with three services: Open WebUI provides the interface, Open Terminal handles terminal commands and access to chosen folders, and SearXNG supplies web search. Open WebUI also brings together tools such as code execution and knowledge bases in the workspace.
The services are separate pieces that need to be connected and configured. Maning reports linking Notion, Obsidian, and LibreOffice through MCP, a way to connect compatible apps and tools to an AI workspace. That gives the setup room to adapt, but it also leaves more configuration and upkeep to the person running it.
Does self-hosting keep AI inference on your computer?
Not by itself. The Docker containers host the applications, while inference—the processing that produces a model’s response—can run locally or through an outside provider, depending on the model connection.
Maning’s configuration could connect Open WebUI to LM Studio through an OpenAI-compatible endpoint, allowing a local model server as one option. In his reported test, a GPU with 8 GB of VRAM ran Qwen3.8-9B-Distill, but prompt processing was too slow for his serious tasks. He then used a paid inference API. That account describes his setup, not a performance result for every computer or model.
The practical distinction is simple: local hosting describes where the workspace services run; it does not determine where every model request is processed. If local inference is a requirement, the model connection matters as much as the Docker setup.
How much access do files and integrations get?
Open Terminal can access host folders that the user mounts into the setup. Maning’s example exposed Documents and Downloads, and he advises sharing only folders the AI may read or modify. Choosing a narrower folder limits what that terminal connection can reach.
A separate option appears in the Open WebUI v0.13.4 release notes: a Tether-based sandbox for Open Terminal. It is described for macOS and supported Windows devices and requires Tether 0.1.13 or later. The sandbox confines terminal processes to one chosen folder; network access is optional, incoming connections are not accepted, and the terminal stops if the sandbox cannot be enforced. Browser and desktop control are disabled while it is active.
Those conditions belong to the separately described sandbox, not to Maning’s Open WebUI 0.13.3 Docker configuration. The sandbox also trades away browser and desktop control while active, so it changes the terminal’s boundaries rather than turning the whole workspace into a different kind of service.
Who may find this setup worthwhile?
This approach suits people comfortable configuring containers, connecting tools, and managing folder access. Maning reports that the setup takes time and that MCP integrations may need attention after service or authentication changes. For someone willing to maintain those pieces, the stack brings search, terminal access, files, and app connections together; anyone who needs local inference should also choose a model route that runs locally.