Caliber

Hand-written CLAUDE.md files go stale the moment you refactor. Your AI agent hallucinates paths that no longer exist, misses new dependencies, and gives advice based on yesterday's architecture. Caliber generates and maintains your AI context files (CLAUDE.md, .cursor/rules/, AGENTS.md, copilot-instructions.md) so they stay accurate as your code evolves — and keeps every agent on your team in sync, whether they use Claude Code, Cursor, Codex, OpenCode, or GitHub Copilot.

Before / After

Most repos start with a hand-written CLAUDE.md and nothing else. Here's what Caliber finds — and fixes:

  Before                                    After /setup-caliber
  ──────────────────────────────            ──────────────────────────────

  Agent Config Score    35 / 100            Agent Config Score    94 / 100
  Grade D                                   Grade A

  FILES & SETUP           6 / 25            FILES & SETUP          24 / 25
  QUALITY                12 / 25            QUALITY                22 / 25
  GROUNDING               7 / 20            GROUNDING              19 / 20
  ACCURACY                5 / 15            ACCURACY               13 / 15
  FRESHNESS               5 / 10            FRESHNESS              10 / 10
  BONUS                   0 / 5             BONUS                   5 / 5

Scoring is deterministic — no LLM, no API calls. It cross-references your config files against your actual project filesystem: do referenced paths exist? Are code blocks present? Is there config drift since your last commit?

caliber score --compare main    # See how your branch changed the score

Get Started

Requires Node.js >= 20.

npx @rely-ai/caliber bootstrap

Then, in your terminal (not the IDE chat), start a Claude Code or Cursor CLI session and type:

/setup-caliber

Your agent detects your stack, generates tailored configs for every platform your team uses, sets up pre-commit hooks, and enables continuous sync — all from inside your normal workflow.

Don't use Claude Code or Cursor? Run caliber init instead — it's the same setup as a CLI wizard. Works with any LLM provider: bring your own Anthropic, OpenAI, or Vertex AI key.

Your code stays on your machine. Bootstrap is 100% local — no LLM calls, no code sent anywhere. Generation uses your own AI subscription or API key. Caliber never sees your code.

Caliber works on Windows with a few notes:

• Run from your terminal (PowerShell, CMD, or Git Bash) — not from inside an IDE chat window. Open a terminal, cd into your project folder, then run npx @rely-ai/caliber bootstrap.
• Git Bash is recommended. Caliber's pre-commit hooks and auto-sync scripts use shell syntax. Git for Windows includes Git Bash, which handles this automatically. If you only use PowerShell, hooks may be skipped silently.
• Cursor Agent CLI: If prompted to install it, download from cursor.com/downloads instead of the curl | bash command shown on macOS/Linux. Then run agent login in your terminal to authenticate.
• One terminal at a time. Avoid running Caliber from multiple terminals simultaneously — this can cause conflicting state and unexpected provider detection.

Audits first, writes second

Caliber never overwrites your existing configs without asking. The workflow mirrors code review:

1. Score — read-only audit of your current setup
2. Propose — generate or improve configs, shown as a diff
3. Review — accept, refine via chat, or decline each change
4. Backup — originals saved to .caliber/backups/ before every write
5. Undo — caliber undo restores everything to its previous state

If your existing config scores 95+, Caliber skips full regeneration and applies targeted fixes to the specific checks that are failing.

How It Works

Bootstrap gives your agent the /setup-caliber skill. Your agent analyzes your project — languages, frameworks, dependencies, architecture — generates configs, and installs hooks. From there, it's a loop:

  npx @rely-ai/caliber bootstrap       ← one-time, 2 seconds
              │
              ▼
  agent runs /setup-caliber             ← agent handles everything
              │
              ▼
  ┌──── configs generated ◄────────────┐
  │           │                        │
  │           ▼                        │
  │     your code evolves              │
  │     (new deps, renamed files,      │
  │      changed architecture)         │
  │           │                        │
  │           ▼                        │
  └──► caliber refresh ──────────────►─┘
       (auto, on every commit)

Pre-commit hooks run the refresh loop automatically. New team members get nudged to bootstrap on their first session.

What It Generates

Claude Code

• CLAUDE.md — Project context, build/test commands, architecture, conventions
• CALIBER_LEARNINGS.md — Patterns learned from your AI coding sessions
• .claude/skills/*/SKILL.md — Reusable skills (OpenSkills format)
• .mcp.json — Auto-discovered MCP server configurations
• .claude/settings.json — Permissions and hooks

Cursor

• .cursor/rules/*.mdc — Modern rules with frontmatter (description, globs, alwaysApply)
• .cursor/skills/*/SKILL.md — Skills for Cursor
• .cursor/mcp.json — MCP server configurations

OpenAI Codex

• AGENTS.md — Project context for Codex
• .agents/skills/*/SKILL.md — Skills for Codex

OpenCode

• AGENTS.md — Project context (shared with Codex when both are targeted)
• .opencode/skills/*/SKILL.md — Skills for OpenCode

GitHub Copilot

• .github/copilot-instructions.md — Project context for Copilot

Key Features

TypeScript, Python, Go, Rust, Java, Ruby, Terraform, and more. Language and framework detection is fully LLM-driven — no hardcoded mappings. Caliber works on any project.

caliber bootstrap auto-detects which agents you have installed. For manual control:

caliber init --agent claude        # Claude Code only
caliber init --agent cursor        # Cursor only
caliber init --agent codex         # Codex only
caliber init --agent opencode        # OpenCode only
caliber init --agent github-copilot  # GitHub Copilot only
caliber init --agent all             # All platforms
caliber init --agent claude,cursor   # Comma-separated

Not happy with the generated output? During review, refine via natural language — describe what you want changed and Caliber iterates until you're satisfied.

Caliber detects the tools your project uses (databases, APIs, services) and auto-configures matching MCP servers for Claude Code and Cursor.

caliber score evaluates your config quality without any LLM calls — purely by cross-referencing config files against your actual project filesystem.

| Category | Points | What it checks | |---|---|---| | Files & Setup | 25 | Config files exist, skills present, MCP servers, cross-platform parity | | Quality | 25 | Code blocks, concise token budget, concrete instructions, structured headings | | Grounding | 20 | Config references actual project directories and files | | Accuracy | 15 | Referenced paths exist on disk, config freshness vs. git history | | Freshness & Safety | 10 | Recently updated, no leaked secrets, permissions configured | | Bonus | 5 | Auto-refresh hooks, AGENTS.md, OpenSkills format |

Every failing check includes structured fix data — when caliber init runs, the LLM receives exactly what's wrong and how to fix it.

Caliber watches your AI coding sessions and learns from them. Hooks capture tool usage, failures, and your corrections — then an LLM distills operational patterns into CALIBER_LEARNINGS.md.

caliber learn install      # Install hooks for Claude Code and Cursor
caliber learn status       # View hook status, event count, and ROI summary
caliber learn finalize     # Manually trigger analysis (auto-runs on session end)
caliber learn remove       # Remove hooks

Learned items are categorized by type — [correction], [gotcha], [fix], [pattern], [env], [convention] — and automatically deduplicated.

Keep configs in sync with your codebase automatically:

| Hook | Trigger | What it does | |---|---|---| | Git pre-commit | Before each commit | Refreshes docs and stages updated files | |