← Hacker News

Show HN: Circus Chief – Claude Code, Codex, and Gemini from Your Phone

Hacker News · deathmonger5000 · June 2, 2026
Circus Chief is a browser-based tool optimized for mobile screens that manages coding agent sessions supporting Claude Code, OpenAI Codex, and Google Gemini. The developer built the application almost entirely from a phone and tablet using AI agents in a "vibe coding" approach, generating 2,184 agent sessions with over 50,000 conversation messages and 310,000 lines of code. The tool enables agents to autonomously operate the application itself, schedule and reschedule work, and execute user-defined commands through configurable prompt templates and worktree isolation.

Detailed Analysis

Circus Chief is a browser-based orchestration tool designed to manage multi-agent coding sessions from mobile and tablet devices, supporting Claude Code, OpenAI Codex, and Google Gemini CLI as underlying agents. Built by developer Ferris Lucas and launched publicly on Hacker News, the tool runs via a simple `npx circuschief` command and is built on a straightforward stack of Vue 3, Express, SQLite, and WebSockets. Its distinguishing architectural feature is that agents themselves can operate the tool — spawning new sessions, scheduling work, interacting with a Kanban board, and executing user-defined commands — creating a recursive layer in which the orchestration system becomes a first-class citizen of the agent's environment rather than merely an external wrapper.

The development story behind Circus Chief is as notable as the product itself. Lucas built the application almost entirely from a phone and tablet, using Claude Code as the primary implementation agent from the project's earliest stages, later supplementing with Codex. The approach was explicitly "vibe coded" — Lucas reviewed high-level plans with agents and iterated on direction, but largely abstained from reviewing individual lines of implementation code. The project then became self-hosting at an early stage, with Circus Chief used to build Circus Chief. The production metrics underscore the scale of this methodology: 2,184 agent sessions since Christmas 2025, dominated heavily by Claude Code at 1,818 sessions versus 363 Codex sessions and only 3 Gemini sessions, generating over 50,000 conversation messages and 310,000 lines of code across 887 tracked files.

The velocity numbers reveal something significant about agent-assisted development at sustained intensity. The repository accumulated 2,410 commits on the main branch and 883 pull requests, with a 96% merge rate. A single peak day on March 21, 2026 produced 76 commits — roughly one every 18 minutes across a 23-hour window. The busiest merge day was New Year's Day 2026, with 28 pull requests shipped. These figures suggest that the primary constraint on agent-driven development shifted away from implementation speed and toward orchestration, review workflow, and session management — precisely the problems Circus Chief is designed to solve.

The tool's design reflects several emerging patterns in the agentic development ecosystem. Automatic rescheduling when usage limits are hit addresses a practical pain point for developers relying on subscription-tier APIs, where rate limits can interrupt long-running autonomous work. Worktree-per-session isolation maps agent sessions cleanly onto Git's worktree model, avoiding the branch collision problems that arise when multiple agents work concurrently on the same repository. The shared canvas feature — a space for planning artifacts that don't belong in the codebase — acknowledges that agent-human collaboration generates a class of intermediate documents that don't fit neatly into traditional source control workflows. Support for third-party providers with Anthropic- or OpenAI-compatible endpoints also reflects the increasingly commoditized nature of model APIs.

Circus Chief sits at the intersection of two converging trends: the maturation of AI coding agents capable of sustaining long autonomous work sessions, and the growing demand for orchestration layers that make those agents manageable outside of desktop development environments. Claude Code's dominance in the session mix — accounting for roughly 83% of all agent sessions — reflects its current strength in agentic coding tasks, while the tool's provider-agnostic architecture positions it to accommodate the continued fragmentation of the frontier model market. The project's own construction history serves as an implicit proof of concept: a substantial, self-hosting production application built primarily through mobile-driven agent orchestration, with the developer deliberately operating at a level of abstraction above individual lines of code.

Read original article →