Skip to content

oneilltomhq/crush

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

114 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

crush

An AI agent runtime that turns voice into action — agents search the web, drive browsers, run shell commands, and manage multi-step workflows while the user directs from a spatial 3D workspace.

What it does

You talk to it. It dispatches specialized agents that do real work: prospecting on LinkedIn, searching job boards, scraping conference sites, building resumes, automating browser sessions. Results appear in a 3D spatial scene — clustered, organized, navigable.

Architecture

Server-authoritative multi-agent system. The browser is a thin rendering client.

Layer What Where
FOH agent Voice interface, task routing, delegation server/agent-server.ts
Worker agents Shell, browser, prospecting, LinkedIn, profiling agents/*.md + server/worker-agent.ts
Skills Domain knowledge with progressive disclosure skills/*/SKILL.md
CLI runner Invoke any agent headlessly server/run-agent.ts
Tool layer Web search, CDP browser automation, shell, filesystem server/pi-tools.ts
3D renderer WebGPU/Three.js spatial scene, SDF text, pane system src/
Voice pipeline Client-side STT/TTS + server relay ADR 005, 006

Agent system

Agent definitions live in agents/ as markdown files with YAML frontmatter (ADR 012). Each specifies a model, tools, skills, and persona. Skills (ADR 013) follow the Agent Skills standardSKILL.md files with references/ directories for deeper context, loaded on-demand.

The FOH agent delegates to typed workers (shell, browser, prospector, linkedin-prospector, linkedin-profiler). Workers load their referenced skills at dispatch time and use read_file to drill into skill references as needed.

# Run any agent from the command line
npx tsx server/run-agent.ts prospector "Find agentic AI contracts in London, posted last 30 days"
npx tsx server/run-agent.ts shell "Update the README and commit"

Key abstractions

  • AgentRunner — Autonomous background worker. Decomposes goals into parallel sub-queries, each with its own LLM conversation.
  • Pane hierarchy — PtyPane (remote shell), BrowserPane (CDP screencast), TextPane (markdown), TerminalPane (Ghostty WASM). Each wraps a Three.js mesh + texture.
  • TaskGraph — Tree of TaskNodes with status lifecycle, driving the spatial scene.
  • grid-scene — The 3D spatial scene. WebGPU rendering, SDF text at 60fps, atmospheric effects.

Running

npm install
npm run dev          # Vite dev server (client)
# Server: npx tsx server/agent-server.ts
# CLI: npx tsx server/run-agent.ts <agent> "<goal>"

Docs

  • adr/ — Architecture decision records (13 and counting)
  • agents/ — Agent definitions
  • skills/ — Domain knowledge skills
  • VISION.md — Design philosophy
  • AGENTS.md — Contributor conventions

About

AI agent runtime — voice-driven spatial workspace where specialized agents search the web, drive browsers, prospect on LinkedIn, and execute multi-step workflows

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors