Hand-written CLAUDE.md files go stale the moment you refactor — Caliber keeps every agent's context accurate as the code changes.
Requires Node.js >= 20.
npx @rely-ai/caliber bootstrapThen, in your terminal (not the IDE chat), start Claude Code or Cursor CLI and type:
/setup-caliber
The agent detects the stack, writes CLAUDE.md / Cursor rules / AGENTS.md / Copilot instructions, and installs the hooks.
No Claude Code or Cursor? caliber init is the same setup as a CLI wizard — Anthropic, OpenAI, MiniMax, or Vertex, your key.
Your code stays on your machine. Bootstrap is local — no LLM calls, no code sent anywhere. Generation uses your seat or your key. Caliber never sees your source.
Windows
- Run from your terminal (PowerShell, CMD, or Git Bash) — not from an IDE chat.
cdinto the project, thennpx @rely-ai/caliber bootstrap. - Git Bash is recommended. Pre-commit hooks and auto-sync scripts use shell syntax. PowerShell-only setups may skip hooks silently.
- Cursor Agent CLI: If prompted to install it, download from cursor.com/downloads instead of the macOS/Linux
curl | bash. Thenagent login. - One terminal at a time. Two Caliber processes in one repo will fight over provider detection.
A typical hand-written CLAUDE.md and nothing else. Scoring is filesystem math — no LLM, no API calls. It checks whether referenced paths exist, whether code blocks are present, and whether the config drifted since the last commit.
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
caliber score --compare main # how this branch moved the score
caliber sync — no LLM, cheap enough for every session start.
caliber sync # mirror into every detected agent
caliber sync --status # what each agent currently holds
caliber sync --dry-run # preview without writing
caliber sync --from cursor # pick the source of truthWhat sync writes, what it refuses to clobber
Caliber Sync
Source: Claude Code — 4 item(s)
Targets: Cursor, Codex, GitHub Copilot
wrote .cursor/skills/deploy/SKILL.md
wrote .cursor/rules/house-style.mdc
wrote .cursor/mcp.json
wrote .agents/skills/deploy/SKILL.md
wrote .github/instructions/deploy.instructions.md
skipped mcp:linear for codex — Codex MCP servers are configured globally
refresh rewrites the prose documents. sync owns the artifacts. They stay in their lanes.
Providers are not symmetric, so sync degrades instead of dropping. Copilot has no skills directory — a skill becomes .github/instructions/*.instructions.md with applyTo from the skill's paths. Codex has no plugin system — a plugin expands into its skills. Anything that cannot be represented is reported with a reason.
Hand edits are never clobbered. Sync hashes every file it writes. A file you edited is a conflict until you pass --force.
Two optional hooks (caliber hooks) keep it continuous:
| Hook | Effect |
|---|---|
Agent sync (SessionStart) |
Every agent starts holding the same skills and rules |
Agent sync (on edit) |
Editing a skill in one agent mirrors it immediately |
The on-edit hook is path-filtered to provider skill/rule directories, so ordinary edits cost nothing.
TypeSafe's Jev scores each tool call and result. Stale ones drop. Everything kept stays byte-for-byte verbatim.
BYOK:
AI_GATEWAY_API_KEYis a Vercel AI Gateway key.TYPESAFE_API_KEYis a TypeSafe key. They authenticate different hosts and are not interchangeable. Caliber does not ship, share, or proxy either.
export CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 # function hooks are early-access, off by default
export AI_GATEWAY_API_KEY=... # Vercel AI Gateway — not a TypeSafe key
# or: export TYPESAFE_API_KEY=... # direct TypeSafe System One
claude plugin marketplace add caliber-ai-org/ai-setup
claude plugin install caliber-jev-compaction@caliberOr vendor it into a repo Caliber already set up:
caliber plugin install --dry-run # print the loadable file tree, write nothing
caliber plugin install # .claude/plugins/ + a local marketplace
caliber plugin list5-minute install → key → live-gate → /compact: FUNCTIONAL_CHECKLIST.md.
Preview the same scoring without installing anything:
caliber compact # report what would be dropped — read-onlyJev Compaction
Messages: 248 -> 201
Chars: 412,880 -> 190,344 (53.9% smaller)
Calls: 96 scored — 41 kept, 23 results dropped, 26 calls dropped, 6 pinned
Keys, fallback, and how to verify the real API
| Key | Host | Model |
|---|---|---|
AI_GATEWAY_API_KEY |
Vercel AI Gateway (/v4/ai/evaluation-model) |
typesafe-ai/jev |
TYPESAFE_API_KEY |
TypeSafe System One (api.typesafe.ai/v1/systemone) |
jev-latest |
Set one in the environment, or enter it when claude plugin install prompts for gatewayApiKey / apiKey.
| Hook | What it does |
|---|---|
session.compact |
Substitutes the verbatim-trimmed transcript for Claude Code's summary |
turn.complete |
Requests compaction once context passes compactAtPercent (60% by default) |
If Jev fails, the key is missing, the transcript will not fit the state budget, or the reduction is below minReductionRatio, the hook logs a fallback and delegates to Claude Code's built-in compaction — a bad Jev day degrades compaction, it does not break the session.
Configuration is plugin userConfig (apiKey, gatewayApiKey, gatewayBaseUrl, keepThreshold, preserveRecentMessages, compactAtPercent, minReductionRatio, maxStateTokens, maxRequestTokens, truncateHeadChars, model) — editable from Claude Code, not a Caliber config file.
Function hooks are an early-access Claude Code surface and stay off unless
CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1is set.caliber plugin installprints the enable steps. Authored against Claude Code 2.1.274; revisit after an upgrade.
The unit suite substitutes the network. To hit the real Jev API:
# Direct TypeSafe System One (a TypeSafe key — not a Vercel key)
TYPESAFE_API_KEY=sk-... npm run e2e:jev
# Vercel AI Gateway — live gate for a Gateway key.
# A Gateway key 401s against api.typesafe.ai. Do not put it in TYPESAFE_API_KEY.
AI_GATEWAY_API_KEY=... npm run e2e:jev:gatewaye2e:jev:gateway skips when the key is unset (no fake pass). HTTP 403 "AI Gateway requires a valid credit card on file" is Vercel billing, not a Caliber bug. They never use a Caliber-hosted or shared secret.
caliber compact is the same scoring as a one-off report — useful before enabling the plugin, or from an agent that is not Claude Code. Below a 25% reduction it tells you compaction is not worth the request. It never rewrites the transcript.
caliber sync also installs a jev-compaction skill into every agent, which is the clearest demo of plugin expansion: Claude Code and Cursor get a skill, Copilot gets an instruction file.
The scoring library is bundled under src/vendor/ (MIT) — no extra install.
Claude Code is the only host with a function-hook plugin that can replace /compact in-session. Cursor agents, Codex, Grok Bot, and similar hosts have no such hook surface today — they get the same Jev scoring through the skill + CLI, and they apply the report themselves.
| Host | Auto /compact toast |
CLI | --write |
|---|---|---|---|
| Claude Code | Yes (caliber-jev-compaction plugin) |
caliber compact (auto-discovers ~/.claude/projects/…) |
No — Claude owns the file; the plugin replaces the session |
| Cursor agents | No | caliber compact --provider cursor --transcript <jsonl> |
No — on-disk JSONL is unofficial, usually has tool_use without ids and no tool_result, so a rewrite would invent fields |
| Grok Bot / Codex / others | No | caliber compact --provider generic --transcript <path> |
Yes — documented caliber.transcript.v1 (or OpenAI-compatible messages JSON) has a proven round-trip |
--provider auto (default) detects the file. --transcript is required when the format is not Claude auto-discover.
# Report (any host). Same BYOK as the plugin.
export AI_GATEWAY_API_KEY=... # Vercel AI Gateway — model typesafe-ai/jev
# or: export TYPESAFE_API_KEY=... # direct TypeSafe — not a Gateway key
caliber compact --provider generic --transcript ./session.json
caliber compact --provider generic --transcript ./session.json --write
caliber compact --provider cursor --transcript ~/.cursor/projects/<slug>/agent-transcripts/<id>/<id>.jsonlPrefer exporting caliber.transcript.v1 from the agent's in-memory conversation (tool results included). Cursor's ~/.cursor/projects/.../agent-transcripts/ file is a thin log: Jev cannot drop results that were never written. The synced jev-compaction skill tells the agent to export, run the CLI, and apply decisions — it does not create a Grok Bot / Cursor toast.
Generic envelope (also accepted: a bare Message[] JSON array, JSONL of those objects, or an OpenAI-compatible messages array):
{
"schema": "caliber.transcript.v1",
"messages": [
{ "role": "user", "text": "Fix the test", "toolUses": [] },
{
"role": "assistant",
"text": "Reading",
"toolUses": [{ "tool_use_id": "call_1", "tool": "Read", "input": { "path": "a.ts" } }]
},
{
"role": "user",
"text": "",
"toolUses": [],
"toolResults": [{ "tool_use_id": "call_1", "text": "file contents" }]
}
]
}Caliber never overwrites your existing configs without asking. The workflow mirrors code review:
- Score — read-only audit of your current setup
- Propose — generate or improve configs, shown as a diff
- Review — accept, refine via chat, or decline each change
- Backup — originals saved to
.caliber/backups/before every write - Undo —
caliber undorestores 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.
Bootstrap gives your agent the /setup-caliber skill. Your agent analyzes the 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.
Claude Code
CLAUDE.md— Project context, build/test commands, architecture, conventionsCALIBER_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.github/instructions/*.instructions.md— Skills and rules degraded into scoped instruction files
Any Codebase
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.
Any AI Tool
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-separatedChat-Based Refinement
Not happy with the generated output? During review, refine via natural language — describe what you want changed and Caliber iterates until you're satisfied.
MCP Server Discovery
Caliber detects the tools your project uses (databases, APIs, services) and auto-configures matching MCP servers for Claude Code and Cursor.
Deterministic Scoring
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.
Session Learning
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 hooksLearned items are categorized by type — [correction], [gotcha], [fix], [pattern], [env], [convention] — and automatically deduplicated.
Auto-Refresh
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 |
| Claude Code session end | End of each session | Runs caliber refresh and updates docs |
| Learning hooks | During each session | Captures events for session learning |
caliber hooks --install # Enable refresh hooks
caliber hooks --remove # Disable refresh hooksThe refresh command analyzes your git diff (committed, staged, and unstaged changes) and updates config files to reflect what changed.
By default the pre-commit hook stages refreshed doc files into the commit in flight. If you prefer to review refreshed docs before committing them (e.g. signed-commit or minimal-diff workflows), disable auto-staging — the refresh still runs, but updated files stay in your working tree:
git config caliber.autostage falseTeam Onboarding
When Caliber is set up in a repo, it automatically nudges new team members to configure it on their machine. A lightweight session hook checks whether the pre-commit hook is installed and prompts setup if not — no manual coordination needed.
Fully Reversible
- Automatic backups — originals saved to
.caliber/backups/before every write - Score regression guard — if a regeneration produces a lower score, changes are auto-reverted
- Full undo —
caliber undorestores everything to its previous state - Clean uninstall —
caliber uninstallremoves everything Caliber added (hooks, generated sections, skills, learnings) while preserving your own content - Dry run — preview changes with
--dry-runbefore applying
| Command | Description |
|---|---|
caliber bootstrap |
Install agent skills — the fastest way to get started |
caliber init |
Full setup wizard — analyze, generate, review, install hooks |
caliber score |
Score config quality (deterministic, no LLM) |
caliber score --compare <ref> |
Compare current score against a git ref |
caliber regenerate |
Re-analyze and regenerate configs (aliases: regen, re) |
caliber refresh |
Update docs based on recent code changes |
caliber sync |
Mirror skills, rules and plugins across every agent (no LLM) |
caliber sync --status |
Show what each agent currently holds |
caliber compact |
Report what Jev compaction would drop (--provider auto/claude/cursor/generic; --write for generic only) |
caliber plugin install |
Install the bundled Jev compaction plugin into this project |
caliber plugin list |
Show bundled plugins and their install state |
caliber skills |
Discover and install community skills |
caliber learn |
Session learning — install hooks, view status, finalize analysis |
caliber hooks |
Manage auto-refresh hooks |
caliber config |
Configure LLM provider, API key, and model |
caliber status |
Show current setup status |
caliber uninstall |
Remove all Caliber resources from a project |
caliber undo |
Revert all changes made by Caliber |
Does it overwrite my existing configs?
No. Caliber shows you a diff of every proposed change. You accept, refine, or decline each one. Originals are backed up automatically.
Does it need an API key?
Bootstrap & scoring: No. Both run 100% locally with no LLM.
Generation (via /setup-caliber or caliber init): Uses your existing Claude Code or Cursor subscription (no API key needed), or bring your own key for Anthropic, OpenAI, MiniMax, or Vertex AI.
Jev compaction (caliber compact / the Claude Code plugin): Yes — your own key as AI_GATEWAY_API_KEY (Vercel AI Gateway) or TYPESAFE_API_KEY (direct TypeSafe). A Vercel key is not a TypeSafe key. Caliber does not provide one.
What's the difference between bootstrap and init?
caliber bootstrap installs agent skills in 2 seconds — your agent then runs /setup-caliber to handle the rest from inside your session. caliber init is the full interactive wizard for users who prefer a CLI-driven setup. Both end up in the same place.
What if I don't like what it generates?
Refine it via chat during review, or decline the changes entirely. If you already accepted, caliber undo restores everything. You can also preview with --dry-run.
Does it work with monorepos?
Yes. Run caliber init from any directory. caliber refresh can update configs across multiple repos when run from a parent directory.
Does it send my code anywhere?
Scoring is fully local. Generation sends a project summary (languages, structure, dependencies — not source code) to whatever LLM provider you configure — the same provider your AI editor already uses. Anonymous usage analytics (command names, durations — no code, no file contents) are collected via PostHog. To opt out:
- Per-run:
caliber --no-traces <command> - Persistent env var:
export CALIBER_TELEMETRY_DISABLED=1
No API key? No problem. Caliber works with your existing AI tool subscription:
| Provider | Setup | Default Model |
|---|---|---|
| Claude Code (your seat) | caliber config → Claude Code |
Inherited from Claude Code |
| Cursor (your seat) | caliber config → Cursor |
Inherited from Cursor |
| Anthropic | export ANTHROPIC_API_KEY=sk-ant-... |
claude-sonnet-4-6 |
| OpenAI | export OPENAI_API_KEY=sk-... |
gpt-5.4-mini |
| MiniMax | export MINIMAX_API_KEY=... |
MiniMax-M3 |
| Vertex AI | export VERTEX_PROJECT_ID=my-project |
claude-sonnet-4-6 |
| Custom endpoint | OPENAI_API_KEY + OPENAI_BASE_URL |
gpt-5.4-mini |
Override the model for any provider: export CALIBER_MODEL=<model-name> or use caliber config.
Caliber uses a two-tier model system — lightweight tasks (classification, scoring) auto-use a faster model, while heavy tasks (generation, refinement) use the default. This keeps costs low and speed high.
Configuration is stored in ~/.caliber/config.json with restricted permissions (0600). API keys are never written to project files.
MiniMax supports OpenAI-compatible and Anthropic-compatible requests in both service regions. Set MINIMAX_BASE_URL or choose a base URL with caliber config:
| Region | OpenAI-compatible base URL | Anthropic-compatible base URL | Documentation |
|---|---|---|---|
| Global | https://api.minimax.io/v1 |
https://api.minimax.io/anthropic |
MiniMax platform docs |
| China | https://api.minimaxi.com/v1 |
https://api.minimaxi.com/anthropic |
MiniMax platform docs |
Vertex AI advanced setup
# Custom region
export VERTEX_PROJECT_ID=my-gcp-project
export VERTEX_REGION=europe-west1
# Service account credentials (inline JSON)
export VERTEX_PROJECT_ID=my-gcp-project
export VERTEX_SA_CREDENTIALS='{"type":"service_account",...}'
# Service account credentials (file path)
export VERTEX_PROJECT_ID=my-gcp-project
export GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.jsonEnvironment variables reference
| Variable | Purpose |
|---|---|
AI_GATEWAY_API_KEY |
Your Vercel AI Gateway key for Jev compaction (not a TypeSafe key; not provided by Caliber) |
TYPESAFE_API_KEY |
Your TypeSafe System One key for Jev compaction (not provided by Caliber) |
ANTHROPIC_API_KEY |
Anthropic API key |
OPENAI_API_KEY |
OpenAI API key |
OPENAI_BASE_URL |
Custom OpenAI-compatible endpoint |
MINIMAX_API_KEY |
MiniMax API key |
MINIMAX_BASE_URL |
MiniMax OpenAI-compatible or Anthropic-compatible base URL |
VERTEX_PROJECT_ID |
GCP project ID for Vertex AI |
VERTEX_REGION |
Vertex AI region (default: us-east5) |
VERTEX_SA_CREDENTIALS |
Service account JSON (inline) |
GOOGLE_APPLICATION_CREDENTIALS |
Service account JSON file path |
CALIBER_USE_CLAUDE_CLI |
Use Claude Code CLI (1 to enable) |
CALIBER_USE_CURSOR_SEAT |
Use Cursor subscription (1 to enable) |
CALIBER_MODEL |
Override model for any provider |
CALIBER_FAST_MODEL |
Override fast model for any provider |
CALIBER_MAX_LEARNINGS |
Cap for CALIBER_LEARNINGS.md bullets (default: 30); evicted entries go to .caliber/learnings-archive.md |
See CONTRIBUTING.md for detailed guidelines.
git clone https://github.com/caliber-ai-org/ai-setup.git
cd caliber
npm install
npm run dev # Watch mode
npm run test # Run tests
npm run build # CompileUses conventional commits — feat: for features, fix: for bug fixes.
After scoring your project, add a badge to your README:
Copy this markdown and replace 94 with your actual score:
Color guide: brightgreen (90+), green (70-89), yellow (40-69), red (<40).
MIT
