caliber-ai-org/ai-setup

What it solves

Caliber provides a unified context layer for coding agents (such as Claude Code, Cursor, and GitHub Copilot). It prevents the common problem of "drift" where the instructions and context files your agents read become outdated as your code changes, and ensures that skills or rules defined for one agent are shared across all others you use.

How it works

Caliber manages a set of context files (like CLAUDE.md or Cursor rules) and uses pre-commit hooks to automatically refresh these files based on your code diffs. It includes a tool called Jev compaction, which optimizes session transcripts by scoring and removing stale tool calls while keeping the rest of the message history verbatim, rather than using lossy summaries.

Who it’s for

Developers who use multiple AI coding agents and want to maintain consistent, up-to-date context and shared rules across their entire development environment.

Highlights

• Automatic Context Refresh: Uses pre-commit hooks to keep agent instructions current with the codebase.
• Shared Skills: Syncs rules, plugins, and MCP servers across different agents deterministically.
• Jev Compaction: A non-summarizing approach to session compaction that drops stale tool calls while preserving verbatim messages.
• Local Audit: The caliber score command audits configuration quality without requiring an LLM.
• Linguistic Agnostic: Supports various providers including Anthropic, OpenAI, MiniMax, and Vertex.