ai-settings

Personal AI agent configuration, harnesses, and infrastructure — one repo to rule all coding assistants.

What is this?

A collection of configurations, skills, tools, and infrastructure for managing AI-powered coding agents across multiple platforms. It provides:

• Agent harnesses — per-agent configuration files (AGENTS.md, opencode.json, etc.) with code style rules, permissions, skills, and model routing.
• Headroom — a Docker Compose deployment of the Headroom API proxy that compresses LLM traffic and reduces costs, with an integrated MCP server.
• Shared rules — common agent instructions (code style, comments, testing) distributed across all harnesses.
• Skills & tools — reusable agent skills, CLI commands, and utility scripts.

Structure

.
├── harness/                  # Agent-specific configuration directories
│   ├── shared/              # Common AGENTS.md, skills, tools, generators
│   ├── opencode/           # OpenCode agent config (opencode.json, plugins, skills)
│   ├── claude/             # Claude Code (placeholder)
│   ├── copilot/            # GitHub Copilot (placeholder)
│   ├── pi/                 # Pi agent configuration
│   ├── qwen/               # Qwen IDE integration
│   └── zed/                # Zed editor (placeholder)
├── headroom/               # Docker Compose: Headroom proxy + MCP server
│   ├── compose.yml         # Default deployment (headroom + MCP)
│   ├── compose-local-router.yml   # Debug mode with local router
│   ├── compose-openrouter.yml     # OpenRouter proxy override
│   ├── compose_proxy.yml          # Slim non-root proxy
│   ├── Dockerfile.mcp      # MCP server image
│   └── .env.base           # Shared environment defaults
└── docs/                   # Planning and specification documents
    └── superpowers/
        ├── plans/
        └── specs/

Quick Start

Headroom (API Proxy + MCP Server)

Deploy the default configuration (headroom proxy + MCP server):

cd headroom
docker-compose -f compose.yml up -d

This starts:

• Headroom proxy on localhost:8787 — compresses LLM traffic at HTTP level
• MCP server on localhost:8788 — exposes headroom_compress, headroom_retrieve, headroom_stats

Debug mode (local router with verbose logging):

docker-compose -f compose.yml -f compose-local-router.yml up -d

OpenRouter proxy (uses your OpenRouter API keys):

docker-compose -f compose.yml -f compose-openrouter.yml up -d

See headroom/README.md for full documentation.

OpenCode

OpenCode is configured in harness/opencode/opencode.json. It includes:

• Providers: oMLX (local models via llama.cpp), OpenRouter
• Models: Qwen 3.6 (27B/35B), Gemma 4, Llama 3.1, Qwen 3.5
• Plugins: opencode-notifier, opencode-handoff, oh-my-opencode-slim
• MCP servers: ai-memory (local), tavily (web search), sequential-thinking

Agent Harnesses

Each directory under harness/ contains configuration for a specific AI agent:

Harness	Purpose	Key Files
shared	Common rules & skills	AGENTS.md, skills/, tools/
opencode	OpenCode agent	opencode.json, AGENTS.md, skills/
pi	Pi agent	agent/, pi-acp/, skills/
qwen	Qwen IDE	ide/
claude	Claude Code	(placeholder)
copilot	GitHub Copilot	(placeholder)
zed	Zed editor	(placeholder)

The shared/AGENTS.md contains project-wide coding standards (functions, comments, tests, dependencies) that apply to all agents.

Features

• Multi-agent support: One repo, many agents. Add a new harness directory with its own config.
• Shared skills: Skills defined in harness/shared/skills/ are available across agents.
• Model routing: Route requests to local models (via llama.cpp/oMLX) or cloud providers (OpenRouter) from a single config.
• Traffic compression: Headroom proxy reduces LLM API costs by compressing requests.
• Cross-session memory: ai-memory provides persistent knowledge across agent sessions.

License

Personal use. All rights reserved.