Back to directory

Caliber

A context layer that keeps CLAUDE.md, Cursor rules, AGENTS.md, and Copilot instructions synchronized to a codebase, writing them from a pre-commit hook and mirroring shared skills, rules, and MCP servers across agents.

Caliber

Caliber is an open-source context layer for coding agents that bootstraps and continuously syncs the setup files each one reads: CLAUDE.md, Cursor rules, AGENTS.md, and Copilot instructions. A pre-commit hook reads the current diff and refreshes those context files, and a no-LLM scoring command checks them against the repository for drift, such as file paths that no longer exist or commands that are no longer real.

Writing a skill, rule, plugin, or MCP server for one agent and running a sync command mirrors it into every other configured agent deterministically, degrading gracefully where a target agent lacks a feature, such as turning a skill into a Copilot instruction file, and never overwriting a file a person edited by hand. A separate Jev-based compaction feature scores each tool call in a session transcript, drops stale ones, and keeps the rest byte-for-byte, as a plugin replacing Claude Code's built-in /compact or as an equivalent CLI command for other agents.

It runs on Node 20 or later and can bootstrap and sync using a Claude Code or Cursor seat or the user's own Anthropic, OpenAI, MiniMax, or Vertex key.

More in Agent skills

See all 581
  • Provenance gate

    github.com

    A deterministic authorization gateway that sits between an AI agent and its tools, tracing where each parameter of a side-effect call came from, and blocking the call or asking a human when the source does not check out.

  • A markdown framework of system rules that restates Isaac Asimov's Four Laws of Robotics as a set of prioritized loss calculations for large-language-model agents to follow, published as a discovery file for agents and crawlers to read.

  • Vitruvius

    github.com

    An engineering research agent that runs a discover, read, synthesize, verify and review loop over technical questions, reading sources directly rather than inferring from titles, and publishing its own scored verifier benchmark, weaknesses included.