Chemaclass/agnostic-ai

agnostic-ai keeps one set of instructions for all your AI coding tools. You write agents, skills, rules, hooks, and MCP server settings once, and agnostic-ai sync writes the files each tool reads.
Each AI tool reads its instructions from different files. Kept by hand, those files fall out of step. agnostic-ai keeps one source in plain Markdown and YAML inside your repository. It needs no account and no service. See Why agnostic-ai.
Set up with a coding agent
Paste this into Claude Code, Codex, Cursor, or another coding agent:
The agent setup guide is the checklist the agent follows.
Quickstart
From your project root, with Node 18 or newer:
Review the planned files. If they match the tools and instructions you want, run:
init --from all creates the project config and imports the tool files it finds. Pick your tools when prompted. sync --plan previews the changes. sync writes the files each tool reads.
Then edit your shared instructions in .agnostic-ai/AGNOSTIC_AI.md and your specs under .agnostic-ai/, and run agnostic-ai sync again. Files such as CLAUDE.md, AGENTS.md, and .cursor/rules/ are generated. Do not edit them by hand.
Next steps:
- Getting started: add your first rule.
- Installation: other ways to install.
- Migration: review an existing setup before you sync.
- Global configuration: personal specs in ~/.agnostic-ai/, written with agnostic-ai sync --global.
Daily commands
Run from the directory containing agnostic-ai.yaml:
When something looks wrong:
The command reference covers import, compare, explain, upgrade, and the other commands.
What you can share
Each tool supports a different set of spec kinds. The capability matrix shows which tools get what, and the target reference shows the exact paths. Supported tools include Claude Code, Codex, Cursor, Gemini CLI, Copilot, and Kiro.
- **Rules, agents, skills, and commands**: instructions in one spec format.
- **Hooks**: checks that run before or after a tool acts. See hooks.
- **MCP servers**: connections to outside services, with credentials read from environment variables. See MCP servers.
- **Capabilities**: what an agent may read, run, or call, mapped to each tool's own names. See capabilities. OpenCode agents keep native permission rule order through import and sync.
- **Model roles**: name a model choice once for every tool. See models.
- **Code review guidance**: instructions for review bots. See reviews.
- **Shared memory**: one project memory that every tool reads and writes. See shared memory.