Introduction

spawn is a terminal UI for orchestrating parallel coding agents — Claude Code, Codex, Grok, and others — from a single board. It adds two things most parallel-agent setups lack: a router that decides which agent and model should handle each task, and a coordination layer that keeps those agents aware of one another while they work. It's open source under the MIT license, built on a Codex-based terminal UI and extended with the spawn-router classifier and a coordination sidecar.

Why spawn

Existing tools for parallel agents hand you infrastructure — parallel worktrees, parallel sessions — but no intelligence about how those agents should cooperate. The moment two tasks overlap, agents collide or go stale: one agent edits files another is mid-way through, and you're left with the familiar "git dirty / another agent edited my files" problem that most tools punt on with blind worktree isolation.

spawn closes that gap with coordination instead of isolation alone:

• A router that decides who does what, so each task lands on the right harness and model tier.
• A coordination sidecar that tracks which agent is touching which files and pushes change summaries into siblings' context before they collide.

The three ideas

• Router-driven dispatch. At every new task, spawn classifies once — task type, complexity, risk, and sub-dimensions — and maps the result through a policy to the right harness and model tier. Inference is local and fast (under 100ms), so there's no per-turn thrash.
• Coordination layer. A live file registry tracks who is editing what right now. On commit, changes are summarized and pushed as one-liners into sibling agents' context, conflicts are headed off at dispatch time, and a diff summary is injected before an agent's next turn.
• TUI board. A board view launches and monitors many parallel sessions at a glance; drill into any session for the full chat, diff, and detail.