Your personal orqestra of specialists. Assembled on demand.

97 specialist agents. 101 skills. One orchestrator that assembles the right team before anyone writes a line of code.

npx orqestra-ai

⭐ Star this repo to get notified when new members join the orqestra.

The problem nobody talks about

You have Claude Code. You have agents. You use them. But every session starts the same way, you describe what you want, Claude responds helpfully, something gets built. Whether it gets built by the right agent, in the right order, with the right context passed between steps: that part is left entirely to you.

That is not a Claude problem. That is a coordination problem. The orqestra solves coordination.

What happens when you type a task

─────────────────────────────────────────────────────────
  you:  "build a SaaS for managing freelance invoices"
─────────────────────────────────────────────────────────
  ORCHESTRATOR

  Pass 1 — which domains does this touch?
    engineering   ✓   API, database, auth, billing logic
    design        ✓   invoice UI, dashboard, empty states
    data          ✓   revenue metrics, invoice analytics
    revops        ✗   no GTM work stated
    people        ✗   no hiring or legal work stated

  Pass 2 — which agents, in what order?
    STAGE 1   system-architect      ← nothing starts without this
    STAGE 2   api-designer          ← parallel
              ux-designer           ← parallel, both wait for stage 1
    STAGE 3   backend-engineer      ← parallel
              frontend-engineer     ← parallel, both wait for stage 2
    STAGE 4   data-analyst          ← waits for stage 3
    STAGE 5   qa-engineer           ← final gate

  every agent receives the output of the previous stage
  every agent knows exactly what to hand off to the next
─────────────────────────────────────────────────────────
  start with @system-architect
─────────────────────────────────────────────────────────

The orchestrator filters by domain cluster first, then selects from a hardcoded registry, it cannot hallucinate agent names. Every pick requires a written justification. If your prompt is too vague to plan against, it asks three questions and stops. No assumptions. No phantom agents. No work done in the wrong order.

The roster

Every agent was built by making Claude inhabit the role first answer as a 15 year veteran, push back on bad decisions, name the mistakes juniors always make, then crystallising that into an agent file. Not a prompt. A perspective.

Engineering — software-engineer frontend-engineer backend-engineer full-stack-engineer mobile-developer system-architect cloud-architect saas-architect solution-architect staff-engineer ai-engineer ml-engineer mlops-engineer data-engineer devops-engineer devsecops-engineer platform-engineer sre security-engineer penetration-tester kubernetes-engineer aws-engineer azure-engineer gcp-engineer embedded-systems-engineer database-administrator api-designer rag-engineer prompt-engineer ai-safety-engineer code-reviewer debugger qa-engineer +more

Design — ux-designer ui-expert design-engineer ux-researcher ux-writer

Data — data-analyst business-analyst bi-specialist ml-analyst analytics-engineer

Business and RevOps — product-manager revops-strategist gtm-expert strategy-expert sales-ops-analyst marketing-ops-analyst crm-specialist customer-success-manager account-executive competitive-intelligence content-strategist ai-product-manager +more

Executive — ceo cfo cmo coo cpo cro cto ciso vp-engineering vp-sales founder

People, Legal, Ops — hr talent-acquisition talent-retention legal compliance-officer procurement it-admin

The one that runs first — orchestrator

101 skills that know when to show up

Skills are dense reference files, written the way a senior expert would write a reference for themselves. They load automatically when an agent needs them. You do not configure them. You do not call them. When backend-engineer is working on your API, api-design-patterns, database-design, error-handling-patterns, and security-best-practices are already there.

A partial list: saas-architecture zero-trust-architecture rag-advanced-patterns llm-integration-patterns agent-design-patterns ai-safety-guardrails vector-database-patterns multimodal-ai-patterns event-driven-architecture service-mesh-patterns kubernetes-patterns terraform-patterns gitops-patterns observability-patterns incident-management sre-practices devsecops feature-flags-patterns cost-engineering finance-modeling executive-frameworks okr-framework gtm-playbook — and 78 more.

What the orchestrator learned from testing

Before this shipped, the orchestrator was tested against real tasks and told to find its own bugs. It found three. Two agents existed in the folder but were missing from the registry — Pass 2 could never select them. It found them, named them, fixed its own registry. A broken critical path reference pointed to a stage that did not exist. It found that too.

An agent planning system that audits itself is not a prompt. It is an opinion about how software should be built.

Install

npx orqestra-ai

Installs to ~/.claude/ globally. Every project you open in Claude Code has access immediately. No config. No API keys. No settings files. Requirements: Claude Code · Node.js 18+

Contributing

Adding an agent takes 10 minutes and requires no coding. Read CONTRIBUTING.md to get started. Want a role that does not exist yet? Open an issue titled [Agent Request] role-name — the community builds it.

Extend it

Every agent lives at ~/.claude/agents/. Every skill lives at ~/.claude/skills/. Open any file, read it, edit it. After adding a new agent, add the name to the cluster registry inside ~/.claude/agents/orchestrator.md so it becomes selectable.

Update

npx orqestra-ai

Same command. Run it again when new agents are released.

Built by Srihari Venkatesan · LinkedIn · Website

There are things in this swarm we have not documented yet. Install it and find out.