Introduction
Welcome to GitCosmo — the open hub for AI agent builders.
GitCosmo is a practical reference for building with AI agents: setting them up, giving them capabilities, and tuning them until they're genuinely useful. Everything works with Claude Code, Codex CLI, Gemini CLI, Cursor, Grok Build, and any other agent that can read markdown and call tools.
Find what you need
| I want to… | Go to |
|---|---|
| Set up my first coding agent | Getting Started |
| Understand what agents actually are | What are AI Agents? |
| Give my agent new capabilities | Understanding MCP + the MCP directory |
| Make my agent follow my project's rules | AGENTS.md Best Practices |
| Write reusable instructions that actually work | Writing Great Skills + the Skills directory |
| Build a repeatable end-to-end workflow | Building Your First Workflow |
| Coordinate multiple agents | Multi-Agent Patterns |
| Fix an agent that's misbehaving | Tuning & Troubleshooting |
| Copy working configs and templates | Examples |
How the pieces fit together
Four things determine how well an agent performs, in rough order of impact:
- The model — pick a strong one; everything else amplifies or wastes it.
- Context — what the agent knows about your project.
AGENTS.mdis the highest-leverage file you can write. - Tools — what the agent can do. MCP servers add filesystem, browser, search, database, and service access.
- Process — plans, self-review, and handoffs. Skills and workflows encode process so you don't re-explain it every session.
Most "my agent is dumb" problems are actually context or process problems. The docs here focus on those two, because that's where you have the most control.
Philosophy
We believe the future of software is built by people directing teams of specialized agents. GitCosmo exists to raise the quality bar for the building blocks — the skills, configs, and patterns — that make those agents reliable.
Everything is open. Everything is designed to be copied and improved.