Skip to content
Work

04 · Developer tooling

Fox's Shelter

A command center for AI agents, projects, and the machines they run on.

Implemented: Working prototype

Agents

1 approval waiting
  • RayCodexWorking
  • Local modelLM StudioAvailable

Repository

Branch
main
Changes
3 files
Tests
Passing
Build
Passing

System

CPU
Memory
Load
Illustration
Status
Working prototype
Environment
  • macOS
  • Local server
  • Git integration
  • Event ingestion API
  • Hardware telemetry
  • Local inference monitoring

Overview

Fox's Shelter is my local command center for agentic development. It brings together what I would otherwise check in several terminals: what each agent is doing, the state of each repository, whether tests pass, and how the machine is coping.

It is a desktop-oriented interface that runs locally, built for the way I actually work: several agents, several projects, one person deciding what ships.

The problem

Directing coding agents across several projects creates a visibility problem. Work happens in sessions you are not watching, on machines with finite memory, and some actions should not happen until a human agrees.

Without one place to look, supervision turns into switching between windows and hoping nothing was missed.

Architecture

  1. Interface

    • Sidebar navigation
    • Work areas
    • Inspector panels
    • Approval inbox
    • Light and dark themes

  2. State

    • Agent sessions
    • Task status
    • Hardware history
    • System status

  3. Ingestion

    • Agent event API
    • Health checks
    • Structured events

  4. Sources

    • Development agents
    • Git repositories
    • Tests and builds
    • Hardware sensors
    • Local models
Agents, repositories, and the machine report into one local service. The interface reads from it.

What exists, and what doesn't

Implemented

  • Agent activity states, session monitoring, task status, and availability
  • Approval queues for actions that should wait for a person
  • Git repository status and working-tree changes
  • Test results and build status
  • CPU, memory, disk, and system load, with historical hardware data
  • Local model availability and inference telemetry
  • System health checks and structured event reporting
  • UX 2.0: sidebar navigation, dedicated work views, inspector panels, system monitoring, activity history, light and dark themes, and consistent design tokens

Explored, not finished

  • Node-based monitoring, where connected machines send structured events and status into one view
  • Integration with ERP AI

My role

  • Product concept and the information architecture of the interface.
  • Direction of the UX 2.0 redesign and its visual hierarchy.
  • Decisions about what agents may do unprompted and what must wait for approval.
  • Implementation was done with AI coding agents working from my specifications. I reviewed the results, set the tests a change had to pass, and decided what was ready.

What I learned

Approval is the feature

The charts are useful. The inbox where an agent's action waits for a yes is what makes it safe to run several agents at once.

Design for a glance

The 2.0 redesign reorganised everything around a sidebar, dedicated work areas, and an inspector, so the first look answers the only question that matters: is anything wrong?

Origins: Gumpun Agent Control

Fox's Shelter grew out of an earlier internal tool, Gumpun Agent Control, which I built while developing Gumpun.

It already carried the core ideas: agent session monitoring, Git status reporting, automated test visibility, local model availability, hardware telemetry, an approval inbox, agent event APIs, and health checks, all running on a local development server on macOS.

Next · 01

Gumpun

One platform for a Muay Thai school's customers, staff, and visa cases.