Building a Centralized, Private Agentic Coding Hub in the Home Lab With Proxmox, Jean, and Local LLMs

Coding with artificial intelligence has fundamentally shifted how developers approach software creation, allowing everyone from seasoned engineers to casual hobbyists to draft scripts, debug errors, and structure complex projects through conversational prompts. However, as agentic coding tools become more advanced, managing them across multiple machines introduces a new set of logistical frustrations. For developers who prefer to keep their workflows entirely self-hosted, setting up repetitive configurations on every new workstation quickly becomes tiresome.

Addressing this friction requires a centralized approach. For developers who already anchor their home infrastructure around hypervisors like Proxmox, deploying a browser-accessible, centralized agentic coding system offers a compelling solution. By combining an open-source workspace manager called Jean with local large language models and secure reverse proxy tooling, it is possible to build an autonomous coding environment that never leaves the local network. While the architecture provides powerful parallel task execution and strict version control, navigating local model behaviors requires careful prompt engineering and active supervision.

What is Jean, and why would you use it?

Jean is an open-source workspace manager developed under the Apache 2.0 license by coolLabs, the same team responsible for the application deployment platform Coolify. Rather than operating as a standalone AI model, Jean serves as a control layer designed to drive existing agent command-line interfaces, including Claude Code, Codex, OpenCode, Cursor, Pi, Grok, Kimi Code, and Antigravity CLI.

The core utility of the platform lies in its execution model. Every coding task assigned to Jean runs in parallel on its own isolated Git worktree and dedicated branch. This architectural choice prevents autonomous agents from accidentally writing modifications directly to a main production branch. Furthermore, a headless server daemon maintains a full user interface accessible via a web browser, secured by a unique access token.

This setup allows a developer to initiate complex coding tasks from a laptop, close the device, and let the home server handle long-running background processes. Users can periodically log into the browser interface to inspect progress, review modifications, or approve stalled tasks. The workspace also features a pre-configured command system known as the Magic Menu, which includes shortcuts for standard actions such as committing changes, code reviews, pull requests, and conflict resolution. These prompts can be customized or expanded within the settings menu to fit specific project requirements.

For developers managing multiple workflows, this centralized approach solves a primary pain point of agentic coding: tracking the simultaneous actions of multiple autonomous agents while maintaining proper Git version history on a single machine rather than scattering repositories across various devices.

Pointing it at a local server makes the whole stack yours and private

While enterprise agentic tools frequently rely on cloud-based subscriptions and external APIs, running an infrastructure stack on a local Proxmox server ensures that sensitive source code remains off the open internet. To achieve complete privacy, developers can route Jean through middleware such as a locally installed OpenCode command-line interface, connecting the dashboard directly to private local language models.

Jean can be deployed across various operating systems, including Windows, macOS, and Linux distributions, though Linux environments integrated into containerized setups have proven particularly effective. Running the platform inside a lightweight Linux Container on a hypervisor like Proxmox isolates the development environment while maintaining persistent uptime for background agent tasks.

Setting up Jean on a Proxmox home lab environment

Deploying agentic development tools often involves complex configuration steps and dependency management. However, installing Jean within a dedicated Debian-based container on Proxmox provides a streamlined path to a centralized workspace. A typical deployment configuration utilizes an unprivileged Debian Linux Container allocated with moderate computing resources, including a few CPU cores, several gigabytes of memory, and modest storage capacity.

Within this container, administrators install Node.js, configure a non-root user account for security purposes, set up OpenCode, and establish secure connections to GitHub and local model providers. Running processes under a non-root user rather than default administrative privileges ensures that autonomous coding agents do not possess unrestricted system access, adding an essential layer of safety to the home lab environment.

To make the workspace easily accessible across a local network without relying on insecure connections, developers frequently integrate a reverse proxy such as Caddy. Configuring Caddy to route traffic to the container enables secure HTTPS communication and assigns a local domain name to the development dashboard.

Despite the relative simplicity of the core installation, auxiliary tooling often introduces minor hurdles. Setting up Debian containers typically requires explicit locale configuration to prevent script execution errors, while Git authentication protocols must be executed under the designated non-root user account rather than as the system administrator. Additionally, managing API authentication tokens for external services like Cloudflare ensures that reverse proxy routing functions reliably. These minor adjustments are standard requirements when deploying multi-dependency software stacks in self-hosted Linux environments.

Local models introduce unique challenges for autonomous workspaces

While self-hosted infrastructure guarantees data privacy and operational independence, relying entirely on local language models introduces variables that do not typically affect cloud-managed frontier models. Software updates across interconnected dependencies can occasionally disrupt system integrations. For instance, simultaneous updates to Jean and OpenCode can sometimes temporarily invalidate user authentication states required to reach local models, requiring administrators to pin specific software versions to maintain operational stability.

Furthermore, autonomous coding agents driven by local models require explicit behavioral boundaries. During testing phases, different local models exhibit distinct operational quirks. Some models attempt to optimize directory structures independently, occasionally misinterpreting complex path hierarchies and initializing unauthorized Git repositories outside the designated workspace boundaries. Others may generate detailed execution plans and confidently report that a task has been successfully completed without actually modifying any files or executing the required tool calls.

These behavioral inconsistencies highlight the importance of operational oversight. Running agents in unsupervised or fully autonomous modes can lead to unexpected directory modifications or false completion claims. Transitioning the workspace into a supervised operational mode, which requires the agent to request explicit confirmation before executing system commands, effectively prevents these unauthorized actions. Additionally, refining system prompts to include explicit worktree paths ensures that local models restrict their file operations strictly to the assigned directory.

Ultimately, deploying Jean within a Proxmox home lab creates a powerful, centralized environment for managing agentic coding workflows. While local language models function efficiently for automated development tasks, they require careful prompt engineering and active supervision to operate reliably within a self-hosted development stack.