AI Cockpit

One interface. Every AI. Runs on your own server.

Setup Guide · User Guide · Configuration Guide · Download

What it is

A local-first web application that orchestrates multiple AI agents — a configurable paid CLI (Claude, Codex, or Qwen Code), a free REST provider (Gemini or Groq), a local LLM server (Ollama, LocalAI, or both at once via Dual mode), cloud Workers (Groq, OpenRouter, Together AI, Z.ai, DeepInfra…), and local CLI tools (ffmpeg, whisper.cpp, pandoc) — from a single browser tab. It's plain PHP + Apache + Bash: no Node.js server, no WebSocket daemon, no persistent background process. It survives Apache restarts and runs on standard shared hosting.

Beyond running them by hand, state a goal and Autopilot Cruise plans it into steps, runs it as a durable, resumable workflow, and hands back the result; underneath, AutoCognition gives your agents a shared local bus where they see, message, and call each other, vote on answers, and route work to the best-fit model. Enter always dispatches; flight PLAN is the only thing that plans.

Your data stays on your machine by default. Cloud model calls use your own API keys, encrypted at rest. Everything above — every model card, local LLMs, workers, tools, RAG, media, and the community Marketplace — is Free, forever, local-first with your own keys. The paid tiers add just three things: Autopilot Cruise, AutoCognition, and Everywhere (remote control) — see Pricing.

Requirements

Runs on Linux, ARM Linux, macOS, and Windows (via WSL2). Every other device connects through the Everywhere relay page on the site instead.

Linux (Mint/Ubuntu-family, CI-verified end-to-end — ARM Linux and macOS are CI-verified too, Windows/Docker aren't yet) with Apache 2 + PHP 8.x
Bash, standard coreutils
Optional: Ollama and/or LocalAI (local models — install both for Dual mode), ffmpeg/whisper.cpp/pandoc (Tools), Claude Code CLI / Gemini CLI (paid agents)
No database required for the local app — everything is flat-file JSON under cockpit_data/

Quick start

# 1. Copy the app into your Apache docroot
sudo cp -r cockpit-bundle/ /var/www/html/cockpit

# 2. Give Apache (www-data) write access to its data folder
sudo chown -R www-data:www-data ~/scripts/cockpit_data

# 3. Open it
http://localhost/cockpit/

# 4. First-run setup walks you through:
#    - creating a local login
#    - pointing at your Ollama and/or LocalAI install (if any)
#    - adding cloud Worker API keys (optional)
#    - authenticating Claude Code / Gemini CLI (optional, paid agents)

Full walkthrough with every step: Setup Guide.

Architecture

Piece	What it does
`index.html`	The entire frontend — single file, vanilla JS, no build step
`paths.php`	Central path resolution — every backend file reads data locations through this one choke point (per-user scoped when a local login is active)
`agent_loop.sh`	tmux queue watcher — dispatches prompts to the active Card 1/2/3 CLI sessions (Claude Code, Gemini CLI, Ollama, or whichever supplier is configured in each slot) and streams responses back via a file queue (no WebSocket needed)
`crud_executor.sh`	Parses `%%CREATE/APPEND/DELETE/READ%%` markers in agent output and performs the file operation on the agent's behalf — the mechanism that lets text-only suppliers (Card 3/Local LLMs, Workers) still touch files
`combos_lib.php` / `combos.php` / `combos_core.php`	Power Combos engine — parallel/sequential/judge multi-model recipes over Ollama and/or LocalAI, dual-mode combos run both servers in parallel at once
`rules.php`	Delegate — keyword/phrase rules that auto-route a matching task to a chosen worker, combo, or Ollama model before the normal agent continues
`cockpit_data/`	All persistent app data — flat JSON, no database

Full feature tour: Features.

License