Questions,
answered.

etyb.ai is a virtual AI engineering company. Instead of a single AI assistant, you get a full team: a CTO-level orchestrator (etyb) that routes every request, 20 domain specialists across the SDLC, and 9 always-on engineering protocols that enforce real engineering culture — TDD, verification, review, plan-execution.

"etyb" is "byte" reversed. We're rethinking what an engineering team looks like from the ground up — starting from the idea that AI should behave like a disciplined team member, not a solo assistant.

etyb.ai is the product and the company. etyb-skills is the open-source foundation that powers it. As of v4 it is one coordinated skill (/etyb) backed by 14 specialists, 9 always-on protocols, 6 vertical architects, and 13 vendor Stacks — all in one repo, MIT-licensed. You can install etyb-skills yourself, inspect how it works, fork it, and run it on any agentskills.io-compliant platform. etyb.ai uses the same system internally.

Teams shipping real software who want their AI agent to behave like a disciplined engineer. Engineers tired of AI that skips tests, rubber-stamps reviews, or chases symptoms. Leaders of regulated codebases (fintech, healthcare, e-commerce) who need traceable decisions and evidence-backed completion. Solo developers who want one expert on demand — every specialist works standalone.

Those are assistants optimized around a code editor. etyb is a virtual engineering company optimized around delivery: a CTO routes work, specialists take ownership, and protocols gate what ships. It is not a single model pretending to know everything — it is a team with defined roles, obligations, and review loops.

Because most requests need more than one discipline. "Build a fintech API with auth" needs fintech-architect + backend-architect + security-engineer, plus TDD, verification, and review protocols. The orchestrator decides the right mix and keeps everyone honest via the protocol layer. Routing is cheap; wrong team composition is expensive.

On Claude Code, the plugin ships shell hooks that fire outside the LLM — no test file before an edit, no review evidence before a commit, merging without a green test run. They surface a warning; they never block, so "hook-enforced" means visibility, not a hard gate. That, plus five model-tiered sub-agents doing the actual work, is why Claude Code is our full-experience platform. On OpenAI Codex, etyb ships 4 lifecycle hooks (prompt guardrails, Bash guards, stop checks) and 5 custom agents (explorer, planner, reviewer, stack-researcher, cartographer) — same advisory model. On Google Antigravity, gates run in model-trusted mode (the model follows the rules rather than the runtime enforcing them).

v4 collapsed the v3 sibling-skill model into one skill with internal references. The reason is trigger discipline: when 30 sibling skills competed for the same prompt, the wrong one would sometimes activate (or two would race), and a user installing only some tiers would silently lose capability. v4 makes /etyb the only trigger surface — it classifies the request and reads the right internal reference under skills/etyb/references/ on demand. You still get full inspectability (every reference is its own README.md), full forkability (delete or swap any reference), and full transparency. The architecture is one skill on the outside, the same 14 + 9 + 6 inside.

Yes. etyb-skills is MIT-licensed and lives at github.com/e-t-y-b/etyb-skills. Transparency is a feature — you should be able to see exactly how your virtual team thinks.

Depends on your agent — two doors, nothing to download or clone on either:

On Claude Code → install the plugin (this is the full team — skill + 5 model-tiered sub-agents + advisory hooks): /plugin marketplace add e-t-y-b/etyb-skills /plugin install etyb@etyb-skills

On any other agent → one npx command: npx skills add e-t-y-b/etyb-skills

The npx route works on every agent that follows the agentskills.io spec (OpenAI Codex, Google Antigravity, Cursor, Cline, Gemini CLI, OpenCode, Warp, Amp, Dexto, Firebender, Kimi Code CLI, Deep Agents, GitHub Copilot — 14+ agents at last count). The CLI fetches the repo itself and asks which agents to wire up, what scope (project or global), and what install method (symlink or copy). Symlink is the default — one source of truth in ~/.skills/, one update propagates everywhere (npx skills update etyb).

Every release ships as a single coordinated skill — the 14 specialists, 9 protocols, 6 verticals, and 13 vendor Stacks come in the box, on every harness.

Two more routes if you want something specific: (1) Claude Code plugin — /plugin marketplace add e-t-y-b/etyb-skills then /plugin install etyb@etyb-skills. Nothing to download — Claude Code fetches the repo itself. Same skill content, plus five sub-agents and five advisory hooks (warn, never block) declared in the plugin's hooks/hooks.json. The full experience. (2) Manual git clone — git clone https://github.com/e-t-y-b/etyb-skills.git, then copy skills/etyb into your harness's skills directory (there's no install script anymore). You get a local working copy you can fork, customize, or pin to a specific commit. All three routes install the same skill; what differs is provenance and whether you get the sub-agents and hooks (Claude Code plugin only).

Use the plugin — on Claude Code it's the difference between the skill and the whole team. From inside a Claude Code session:

/plugin marketplace add e-t-y-b/etyb-skills /plugin install etyb@etyb-skills

Nothing to download; Claude Code fetches the repo itself. The plugin installs the skill PLUS five sub-agents (model-tiered — cheap models for cheap work, never above your own session model) and five advisory hooks that warn — no test file before an edit, no review evidence before a commit — but never block.

npx skills add e-t-y-b/etyb-skills also works on Claude Code, but it installs the skill only — the sub-agents and hooks live in the plugin manifest, so npx can't carry them. /etyb still runs with the full knowledge tree; its heavy work just runs instruction-guided instead of through the tuned sub-agents. If you installed via npx first, npx skills remove etyb, then install the plugin — running both makes /etyb discoverable twice.

One command:

npx skills add e-t-y-b/etyb-skills

That installs the skill into .agents/skills/. The repo also ships committed .codex/ runtime assets — config.toml, lifecycle hooks (prompt guardrails, Bash guards, stop checks), and five custom agents (explorer, planner, reviewer, stack-researcher, cartographer) generated from the same source as the Claude Code agents — so a git clone picks them up automatically with no separate runtime-install step. Edit-before-test remains a model-trusted gap. See docs/installation.md for details.

Same npx skills add e-t-y-b/etyb-skills command — it installs into .agents/skills/. Antigravity additionally supports an optional ADK (Agent Development Kit) integration for parallel sub-agent tracks — see skills/etyb/adapters/antigravity/adk-integration.md, a documented future path rather than a shipped runtime today. Gates run in model-trusted mode.

One command, mirroring whichever route you used to install:

Universal (npx skills) → npx skills update etyb. With the default symlink install, this updates every agent on your machine at once.

Claude Code plugin → update from the /plugin menu, or claude plugin update etyb@etyb-skills. The marketplace tracks releases and re-wires hooks and agents automatically.

Manual clone → cd etyb-skills && git pull, then re-copy skills/etyb to your target directory. There's no update script anymore — it was retired along with install.sh in v5.0.0.

ETYB ships a release roughly weekly — vendor Stack pages get currency-stamped updates, so re-running the update is also how you keep platform knowledge current. The signature footer on every Tier 1-4 response links the changelog at etyb.ai/changelog.

Not since v4. The v3.x bundles (full, process-protocols, core-team, verticals) and the --bundle / --skills flags were removed when the architecture collapsed into a single coordinated skill. Every install copies the full /etyb — 14 specialists, 9 protocols, 6 verticals, 13 vendor Stacks — because their on-disk size is small (everything is markdown) and trying to install partial sets caused real bugs in v3 (some references pointed at missing siblings). If you genuinely need a tier-style install, the v3.x release line still exists.

You don't install one — you invoke /etyb with a narrow request and only the relevant specialist's reference loads. "Review this React component" loads frontend-architect. "Audit this Apex class for bulkification" loads backend-architect plus the Salesforce stack overlay. The orchestrator handles the routing; you don't need to name the specialist. If you want to force a specific lane, just say so — "have the security-engineer look at this" works.

Just describe the work in natural language. The orchestrator listens for intent that spans multiple disciplines, is ambiguous, or needs end-to-end planning — phrases like "build me…", "design and implement…", "ship a production-ready…". If your request is narrow ("review this function"), the relevant specialist takes it directly without orchestration overhead.

No. Skills auto-activate from your request. "Build a HIPAA-compliant telehealth backend" triggers healthcare-architect + backend-architect + security-engineer automatically, along with TDD, plan-execution, and verification protocols. You can call a skill explicitly (e.g. "run the review-protocol on this PR") when you want to force a specific workflow.

They are always on. TDD fires whenever implementation begins. Verification fires before any "done" claim. Review fires on PRs. Debugging fires when tests repeatedly fail. Plan-execution drives the task loop. You do not opt in — you opt out if a protocol does not apply (and etyb logs the reason).

Lead with intent, not with instructions. "Build a multi-tenant SaaS with usage-based billing" is better than "write a file at path X". Let the orchestrator decompose the work, propose a plan, and wait for your sign-off before executing. If the request is ambiguous or greenfield, ask etyb to brainstorm first — the brainstorm-protocol produces a design brief before any code is written.

Every completion triggers the verification-protocol: five explicit questions (does it compile, do tests pass, does it solve the stated problem, what edge cases were considered, what could go wrong) plus a structured completion report. On Claude Code, the TDD hook and branch-safety hook physically refuse to close the task if gates fail. No "works on my machine" claims of done.

Any request where discipline matters more than speed: greenfield projects, regulated domains, production-ready shipping, end-to-end features. Narrow tweaks ("change the button color") work too, but you're using a small fraction of the system. The system earns its weight when the request spans planning, implementation, testing, review, and deployment.

14 always-available domain experts: backend-architect, frontend-architect, mobile-architect, system-architect, database-architect, security-engineer, devops-engineer, sre-engineer, qa-engineer, ai-ml-engineer, project-planner, research-analyst, code-reviewer, technical-writer. Each is a standalone skill with 100+ deep references.

6 vertical specialists loaded when the domain applies: saas-architect (multi-tenancy, billing), fintech-architect (ledgers, payments, PCI/AML/KYC), healthcare-architect (HIPAA, FHIR, EHR), e-commerce-architect (catalog, checkout, fulfillment), real-time-architect (WebSockets, CRDTs, streaming), social-platform-architect (feeds, ranking, moderation).

TDD, plan-execution, verification, debugging, review, brainstorm, git-workflow, subagent, and skill-evolution. They enforce process discipline that no individual skill owns: no code without a failing test, no claims without evidence, no merges without green tests, no skill changes without a failing eval first.

13 vendor Stacks, all in the same repo: AWS, GCP, Azure, Cloudflare, Vercel, Supabase, Firebase, Stripe, Salesforce, Anthropic Claude, OpenAI, Expo, Observability. Each Stack is a folder under stacks/<vendor>/ with a slim SKILL.md trigger pointer, an index.md briefing, per-product canonical pages, and per-role composed views. Every page carries last_verified_on + drift_risk + authoritative_url frontmatter — when a page is older than its drift threshold, /etyb either discloses the staleness or fetches the authoritative URL for high-stakes claims. See /stacks for the full catalog and /stacks/salesforce for the deepest example.

Check the skill's SKILL.md triggers list. Every skill has an explicit list of trigger phrases; if your request does not match any, the skill will not auto-activate. You can force it by naming the skill: "run fintech-architect on this ledger design." If triggers feel wrong, that is a skill-quality bug — file an issue at github.com/e-t-y-b/etyb-skills.

github.com/e-t-y-b/etyb-skills/issues for skill bugs, missing triggers, or new-skill requests. For product issues on etyb.ai (this site), email hello@etyb.ai.

Yes. Read skill-evolution-protocol — it is the meta-protocol for skill creation. Every new or modified skill needs a failing eval first (TDD for documentation), then content that makes the eval pass. PRs welcome at github.com/e-t-y-b/etyb-skills.

Email hello@etyb.ai. We work with teams on high-stakes or regulated codebases where the protocol discipline matters most.