Vendor-neutral AI tooling engine

AI coding standards,
shipped like a dependency.

Author rules, agents, and skills once. Bluetemberg compiles them into native config for six AI tools — routed by each engineer's role and the stack's exact version, then verified like a real dependency.

npm version MIT license Node 20+ SHA-512 verified

Get started

$npx bluetemberg init

?Team profile:

Frontend — UI, design systems, accessibility
Backend — APIs, databases, auth, services
Full-stack — Frontend + backend combined
DevOps / Platform — CI/CD, containers, infrastructure-as-code
Pure Infrastructure — Ansible, Kubernetes, Terraform — no application code
AI / Agentic Workflow — Context engineering, agent memory, sub-agent design for LLM-heavy projects
Design Engineer — Anti-stock UI craft, design-to-code, visual direction & critique
Custom — Pick everything individually

Requires Node.js 20+.  Headless: --non-interactive --profile backend.  All commands & flags →

How it works

One source of truth. No copy-paste between tools.

01 / AUTHOR Write once Rules, agents, and skills as plain Markdown in llm/. Vendor-neutral — no tool-specific frontmatter.
02 / SYNC Compile everywhere bluetemberg sync emits native config for every tool — correct frontmatter, correct paths, per platform.
03 / SHIP Onboard instantly Publish packs to npm with semver, SHA-512, and registry signatures. Teammates install — verified on arrival — or get them zero-setup via a Claude Code plugin.

Version-aware routing

Two axes, not one. Role decides who gets a rule. Stack decides which version of it.

  • By role Profiles match content to an engineer's work — frontend, backend, infra, design. Each teammate gets the standards that apply to them, and none that don't.
  • By stack & version Tag a rule with a version range; sync delivers it only where it's correct — resolving each stack's version from your config, lockfile, and dependencies. A Payload 2 rule never reaches a Payload 3 repo.
  • Queryable bluetemberg detect reports the stacks and versions it found; coverage checks whether version-correct guidance exists. Both have --json twins — and bluetemberg mcp serve exposes them to any agent over MCP. Read-only — it answers, it never writes.

Humans + agents

Your team writes the standards. So can their agents.

  • Humans author Rules, agents, and skills are plain Markdown in llm/ — owned by the team, reviewed in PRs, versioned with your code. Run the wizard, pick a profile, done.
  • Agents operate Every command has a headless twin. init --non-interactive skips the wizard, sync --check fails CI on drift, and --help --json exposes the whole catalog. No keyboard required.
  • Agents author too Skills like create-rule, create-skill, and create-pack teach your AI to write new standards in the exact format — then sync ships them to every tool. AI config, authored by AI.

Tool coverage

Six tools, native formats.

Rules Agents Skills
Cursor ● ● ●
Claude Code ● ● ●
Copilot ● ● ●
Gemini CLI ● – –
Windsurf ● – ●
OpenAI Codex ● ● ●

Supply chain

Packs are dependencies. So they're verified like dependencies.

  • Registry signatures Packs from the npm registry have their ECDSA registry signature verified on every install — and re-checked any time with bluetemberg verify. Tampered or unsigned packages are refused — the same cryptographic provenance check npm itself runs.
  • SHA-512 integrity Every pack is hashed and checked against npm registry metadata. Install is refused if the hash is missing or doesn't match.
  • Registry host pinning Tarballs must download from the same host as the registry. A compromised registry response can't redirect a download to an attacker's host.
  • Size caps 50 MB compressed, 100 MB uncompressed. A malicious pack can't fill your disk or bomb the extractor.
  • Path-traversal & symlinks Every tarball entry is filtered — .. segments, symlinks, and hardlinks. Nothing can write outside the pack's own directory.
  • Lockfile Exact versions, resolved URLs, and integrity hashes pinned in llm/packages-lock.json. Every install resolves to the same bytes — the same model npm uses for package-lock.json.

Standards with receipts

Most AI rules are a model's best guess. Ours are checked against the source.

  • Primary sources Every claim past plain convention is verified against official docs, IETF RFCs, and vendor advisories — PostgreSQL, RFC 9110, OWASP, CISA. Sourced from the spec, not forum folklore.
  • Default-refute An independent agent tries to disprove each claim; only the ones that survive a majority of checks ship. Overstated guidance gets downgraded, not waved through.
  • Public receipts Every verdict and source is on the Research page — confirmed, or flagged "nuanced" and reworded. Including the claims we walked back.

bluetemberg vs. a shared AGENTS.md

A file in a repo isn't infrastructure.

  Shared AGENTS.md Bluetemberg
Versioning git history semver ranges + lockfile
Integrity none SHA-512, ECDSA-signed, host-pinned
Provenance whatever the model knows verified vs. primary sources
Per-role config manual copy-paste profiles, role-matched defaults
Stack targeting one file, every version version-gated at sync
Multi-tool one file, one format 6 tools, native formats
Onboarding clone & copy files install, or zero-setup plugin
Versioning git history semver ranges + lockfile
Integrity none SHA-512, ECDSA-signed, host-pinned
Provenance whatever the model knows verified vs. primary sources
Per-role config manual copy-paste profiles, role-matched defaults
Stack targeting one file, every version version-gated at sync
Multi-tool one file, one format 6 tools, native formats
Onboarding clone & copy files install, or zero-setup plugin

Explore

Browse packs ↗

Every published rule, agent, skill, and guardrail — searchable and filterable by role.

Read the docs ↗

Installation, commands, configuration, the security model, and writing your own packs.

View on GitHub ↗

Source, issues, and contribution guide. MIT licensed, Node 20+.