ai-setup
caliber-ai-org · caliber-ai-org/ai-setup
Đồng bộ hóa cấu hình, kỹ năng và tệp ngữ cảnh của tác nhân AI trên Claude Code, Cursor và Codex, với các tính năng làm mới ngữ cảnh liên tục và nén.
Continuously sync your AI setups with one command. Codebase tailor suited agent skills, MCPs and config files for Claude Code, Cursor, and Codex.
Xem trước
Bạn có thể làm gì với nó
- Giữ các kỹ năng Claude Code và quy tắc Cursor được đồng bộ hóa trên nhiều môi trường phát triển.
- Tự động làm mới các tệp ngữ cảnh của tác nhân AI dựa trên các thay đổi mã bằng cách sử dụng pre-commit hook.
- Củng cố và nén bộ nhớ hoặc ngữ cảnh của tác nhân bằng cách sử dụng nén Jev để tăng hiệu quả.
Kiểm tra an toàn
Không thấy dấu hiệu rủi ro trong các file đã quét.
Quét tự động các file trong repo. Luôn xem lại code trước khi chạy.
Trích từ README
The open-source context layer for coding agents. Claude Code · Cursor · Codex · OpenCode · GitHub Copilot
Start
npx @rely-ai/caliber bootstrap
Then, in Claude Code or the Cursor CLI, run /setup-caliber. Your agent writes CLAUDE.md, Cursor rules, AGENTS.md and Copilot instructions, shows you the diff, and installs the hooks that keep them current. No agent handy? caliber init does the same as a wizard.
Node 20+. Runs on your Claude Code or Cursor seat, or your own Anthropic, OpenAI, MiniMax or Vertex key. Bootstrap, scoring and sync run fully local.
Fresh
A pre-commit hook reads your diff and refreshes the context files your agents read. caliber score checks them against the repo, with no LLM: do the paths exist, are the commands real, has anything drifted.
caliber score --compare main
Shared
Write a skill, rule, plugin or MCP server in one agent and caliber sync puts it in every other one. It's deterministic, degrades where an agent lacks a feature (a skill becomes a Copilot instruction file), and never overwrites a file you edited by hand.
caliber sync --status
Lean: Jev compaction new
/compact summarizes, and summaries forget the file path, the error and the constraint you gave an hour ago. TypeSafe's Jev scores every tool call, drops the stale ones, and keeps everything else byte-for-byte.
Claude Code: a plugin that replaces /compact.
export CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1
export AI_GATEWAY_API_KEY=... # or TYPESAFE_API_KEY
claude plugin marketplace add caliber-ai-org/ai-setup
claude plugin install caliber-jev-compaction@caliber
Cursor, Codex and any other agent: the same scoring from the CLI.
caliber compact --provider cursor --transcript <jsonl>
caliber compact --provider generic --transcript ./session.json --write
Bring your own Vercel AI Gateway or TypeSafe key. If Jev is unavailable, Claude Code falls back to its built-in compaction. Setup and verification: FUNCTIONAL_CHECKLIST.md.
Commands
| Command | |
|---|---|
caliber bootstrap / init | Set up (agent skill or CLI wizard) |
caliber score | Audit config quality, no LLM |
caliber refresh | Update configs from recent changes |
caliber sync | Mirror skills, rules, plugins and MCP across agents |
caliber compact | Jev compaction report (--write for generic transcripts) |
caliber plugin install | Install the Jev plugin into this project |
caliber learn | Learn patterns from your sessions into CALIBER_LEARNINGS.md |
caliber hooks | Manage refresh and sync hooks |
caliber config | Choose provider, key and model |
caliber undo / uninstall | Revert changes / remove everything Caliber added |
Every write is shown as a diff first and backed up to .caliber/backups/. Anonymous usage analytics (command names, never code) can be turned off with CALIBER_TELEMETRY_DISABLED=1.
Contributing
git clone https://github.com/caliber-ai-org/ai-setup.git && cd ai-setup
npm install && npm run test
See CONTRIBUTING.md. MIT licensed.