Caliber — AI Agent Config Sync
Keep your AI setups in sync with one command — generate and maintain context files for Claude Code, Cursor, Codex, and more.
Evidence: CLI requires user confirmation before writes (diff review), provides backups and undo, API keys stored in ~/.caliber/config.json with 0600 permissions, and claims no code is sent. Deductions: dependencies use caret ranges (^) posing supply chain risk; telemetry enabled by default requiring opt-out; publisher unverified, source attribution unclear.
Evidence: README and package.json descriptions align, CI configured for multi-platform tests. Deductions: dependency version ranges are broad, potentially causing non-reproducible builds; failure messages not detailed in docs, relying on CLI output.
Evidence: Supports multiple AI tools and languages, provides multiple installation methods (bootstrap/init), and includes Windows-specific notes. Deductions: capability boundaries not explicit (e.g., MCP discovery scope), triggers like /setup-caliber depend on user action, environment fit not covering all scenarios.
Evidence: README well-structured with install notes, FAQ, command list, MIT license clear. Deductions: no explicit changelog, naming stability not guaranteed (frequent version updates), known limitations not listed, maintenance responsibility unclear (publisher unverified).
Evidence: Output is usable config files, provides scoring and auto-fix, high marginal value (saves manual maintenance). Deductions: cost-benefit not quantified, no real-world usage cases or performance data.
Evidence: README claims (e.g., local operation, deterministic scoring) align with code structure, but no independent verification. Deductions: lack of traceable test results, insufficient cross-source corroboration, facts vs inferences not clearly separated.
- Dependencies are not pinned, posing supply chain risk; consider locking versions.
- Telemetry is enabled by default, requiring opt-out; be aware of privacy.
- Publisher is unverified, source attribution unclear; evaluate before use.
What does this agent do, and when should you use it?
Caliber is a CLI tool (npm package @rely-ai/caliber) that solves the problem of stale AI agent context files (CLAUDE.md, .cursor/rules/, AGENTS.md) by automatically generating and maintaining them. It uses deterministic scoring (caliber score) to audit existing configs, generates tailored context files, MCP configurations, skills (OpenSkills format), and hooks for multiple platforms. Caliber supports various LLM providers (Anthropic, OpenAI, MiniMax, Vertex AI) but can leverage existing Claude Code or Cursor subscriptions without extra API keys. Installation is via npx, and the initial setup is handled by the agent itself via /setup-caliber. All writes are diffed, backed up, and reversible.
Caliber provides two primary commands: caliber bootstrap (installs agent skills for fastest start) and caliber init (full interactive wizard). caliber score performs a deterministic audit, cross-referencing config files against the actual filesystem, and outputs a score with structured fix data. caliber refresh analyzes git diff (committed, staged, and unstaged) and updates config files to reflect code changes. caliber learn installs hooks to capture AI session events and distill patterns into CALIBER_LEARNINGS.md. caliber hooks manages git pre-commit and session-end hooks for auto-refresh. caliber undo and caliber uninstall provide full reversibility. Generated files include CLAUDE.md, CALIBER_LEARNINGS.md, .claude/skills/*/SKILL.md, .mcp.json, .cursor/rules/*.mdc, AGENTS.md, .github/copilot-instructions.md, and more.
- A developer who just refactored a project wants CLAUDE.md and Cursor rules updated automatically, so the AI doesn't hallucinate outdated paths.
- A team wants to keep AI configs consistent across Claude Code, Cursor, Codex, OpenCode, and GitHub Copilot, ensuring every member's agent behaves the same.
- A user wants to assess the quality of their current AI setup before making changes, using caliber score --compare main to see how a branch impacts the score.
- A developer who already has a Claude Code or Cursor subscription wants to generate configs without paying for additional API keys.
- A developer wants configs to refresh automatically on every commit via pre-commit hooks, reducing manual maintenance.
- A team working in a monorepo wants to run caliber refresh from a parent directory to update configs across multiple repositories at once.
What are this agent's strengths and limitations?
- Deterministic scoring (no LLM) provides consistent, auditable config quality checks, with git branch comparison.
- Bootstrap and scoring are fully local and don't send code; generation uses your own AI subscription or API key.
- Automatic backups, score regression guard, full undo, and clean uninstall ensure safety and reversibility.
- Supports multiple platforms (Claude Code, Cursor, Codex, OpenCode, GitHub Copilot) to unify team configs.
- Automatic MCP server discovery and skills in OpenSkills format enhance agent capabilities.
- Generation requires network and third-party APIs (unless using Claude Code or Cursor subscription), and sends a project summary to the LLM provider.
- Generates many files, potentially increasing repo complexity; requires Node.js >= 20.
- Windows support has caveats: Git Bash recommended, avoid multiple terminals, hooks may be silently skipped in PowerShell.
- Core operation depends on specific agent environments (e.g., /setup-caliber for Claude Code), requiring extra configuration for cross-platform use.
- Scoring criteria may favor certain config styles and may not suit all project structures.
How do you install or deploy this agent?
Requires Node.js >= 20. Run in your project root: npx @rely-ai/caliber bootstrap. No API key is needed for bootstrap or scoring (fully local). Alternatively, use caliber init for the full wizard. For generation, you can configure an LLM provider: export ANTHROPIC_API_KEY, OPENAI_API_KEY, or MINIMAX_API_KEY, or use your existing Claude Code or Cursor subscription.
How do you use this agent?
Run npx @rely-ai/caliber bootstrap. Then, in your terminal (not IDE chat), start a Claude Code or Cursor CLI session and type /setup-caliber. The agent will analyze your codebase, generate configs, install hooks, and enable continuous sync. If you don't use Claude Code or Cursor, run caliber init as a CLI wizard. Common commands: caliber score to see your score, caliber refresh to update docs, caliber undo to revert changes, caliber hooks --install to enable auto-refresh hooks. All writes are diffed and backed up first.