What Happened
A developer known as moridinamael has released PlateSpinner, an open-source (MIT licensed) local web app that manages multiple AI coding agents through a kanban board interface. You point it at a project directory, describe what you want built, and it spawns headless sessions of Claude Code, Codex, or Gemini CLI to generate tasks, plan implementations, and write code.
The workflow has three phases. Propose runs a read-only AI session that analyzes your codebase and returns structured task proposals with titles, descriptions, rationale, and effort estimates. Plan spawns another read-only session producing concrete implementation details - specific files to modify, functions to create, tests to add. Execute launches a full-access session where the agent writes code, runs tests, and commits changes. Each task gets its own git branch like kanban/task-a1b2c3d4-add-dark-mode.
The supported models span the major players: Claude Opus 4.6 at $15/$75 per million tokens (the default), Gemini 3.1 Pro at $2.50/$15, and GPT-5.3 Codex at $5/$15. PlateSpinner calls these through their respective CLI tools rather than cloud APIs directly.
The most ambitious feature is "autoclicker mode" - an autonomous loop where an AI judge analyzes project state and decides whether to propose, plan, or execute next. It supports up to 10 concurrent agents with configurable budget caps and automatic backoff to prevent runaway costs.
The project is built with React, Vite, and Express, uses WebSocket for real-time streaming, and stores state in a JSON file. It has 77 commits and 2 contributors (the developer and Claude, listed as a contributor). It is early - 1 star, no releases published yet.
Why It Matters
AI coding assistants have been single-player tools. You open Cursor or Claude Code, work on one task, get help, move on. PlateSpinner approaches the problem differently: what if you could manage multiple AI agents working on different parts of your codebase simultaneously, the way a tech lead manages a team?
The three-phase workflow is smart design. Read-only propose and plan phases let you review what the AI intends to do before giving it write access. Branch-per-task isolation means a bad execution does not trash your working tree. The diff viewer lets you review every change before pushing. These are the guardrails that make autonomous coding less terrifying.
The cost differences between models also become meaningful at scale. Running 10 concurrent Gemini sessions costs roughly one-sixth of the equivalent Claude Opus sessions. PlateSpinner makes model selection a practical, per-task decision rather than a one-time setup choice.
Our Take
PlateSpinner is the kind of tool that shows where AI-assisted development is heading - from pair programming to project management. The kanban metaphor is a natural fit. Developers already think in terms of task boards. Attaching an AI agent to each card is a logical extension.
The autoclicker mode is both the most interesting and most dangerous feature. Letting an AI decide what to work on next, with up to 10 agents running in parallel, could burn through API credits fast. The budget caps are essential, not optional. If you try this, start with one agent and watch what it does before scaling up.
What I like is the architecture choice of using CLI tools rather than API calls directly. This means PlateSpinner works with whatever authentication and rate limiting your existing Claude Code or Codex setup already has. No separate API keys to manage, no new billing relationships.
The 1-star, 77-commit state tells you this is very early. The contributor list including "Claude" as a named contributor is fitting for the era we are in. If you are already using Claude Code or Codex from the command line and want to run multiple tasks in parallel with a visual interface, PlateSpinner is worth trying. If you are looking for something production-stable, give it a few months.
For teams evaluating AI coding workflows, this is a preview of where tools like Cursor and Claude Code are likely headed - multi-agent orchestration with human oversight at the review stage, not the typing stage.