agentstack

The Rails for AI agencies. Multi-tenant by default.

If you build AI agents for more than one client, you've solved the same five problems N times: how to keep voice consistent, how to scope memory and RAG, how to enforce review, how to ship without breaking other clients, how to measure quality. agentstack solves them once.

agentstack is an opinionated workflow for building production AI agents at scale across multiple clients. It turns Claude Code into your agency's playbook: an architect who scaffolds the company, a designer who locks the voice, an engineer who writes the state machine, a lawyer who reviews the disclosure, and a release manager who ships the PR. Eight slash commands. Four review lenses. Eight Iron Laws.

Built on Mastra. MIT.

---

What you can do in 12 minutes

# Scaffold a new agency repo
$ pnpm create agentstack my-agency
$ cd my-agency

# Onboard a new client
$ claude
> /agentstack-new-company acme-creators
# [interview, CONTEXT.md + INTEGRATIONS.md filled]

# Build a production-ready agent for them
> /agentstack-autoplan setter for acme-creators
# [runs four-lens review → scaffolds agent → drafts tools, role, scorers]
# [surfaces three taste calls; you decide]

# Verify against golden inputs
> /agentstack-qa acme-creators/ig-setter

# Ship
> /agentstack-ship-agent acme-creators/ig-setter
# [creates PR with review comments from each lens]

12 minutes from "we need a setter for this client" to "PR with four-lens review attached."

---

Install — 30 seconds

Requirements: Claude Code, Node.js 22.13+.

npx agentstack-skills

That fetches the latest released tag of this repo from GitHub and copies the eight skills + root dispatcher to ~/.claude/skills/agentstack-*. Restart Claude Code or open a new project.

To pin to a specific version, install from main, or uninstall — see the agentstack-skills readme.

Alternative: install from a local clone (for contributing to agentstack itself)

git clone https://github.com/agentpilled/agentstack.git ~/agentstack
cd ~/agentstack && bash setup/install.sh

Verify:

> /agentstack-validate

Next: follow the getting-started walkthrough — 6 steps, ~10 minutes, from a fresh clone to your first running agent.

---

Skills

Slash command	What it does
/agentstack-new-company	Interview + scaffold a new client workspace (CONTEXT, INTEGRATIONS, env)
/agentstack-new-agent	Interview + scaffold an agent for an existing company
/agentstack-validate	Run the 8 Iron Laws against the repo
/agentstack-plan-review	Four-lens review of an agent spec
/agentstack-autoplan	The killer. Spec → 4 reviews → scaffold → drafts. Auto-decides via 6 principles
/agentstack-qa	Replay golden inputs, score, fix iteratively
/agentstack-schedule	Put a runner on a recurring schedule (launchd / systemd) with .env loading + locking + window of operation
/agentstack-learn	Extract patterns post-build, propose updates to SKILL/IRON-LAWS
/agentstack-ship-agent	Validate → review → register → tests → PR

---

Foundations

Read in this order:

1. ETHOS.md — what we believe and what we reject
2. IRON-LAWS.md — eight invariants nothing ships without
3. DECISION-PRINCIPLES.md — six principles /agentstack-autoplan uses to auto-resolve
4. lenses/ — four review personalities (operator, lead-user, stakeholder, compliance)
5. SKILL.md — the six conventions every agentstack repo follows
6. ARCHITECTURE.md — repo shape, the agent journey, package model

Patterns

Field-extracted patterns (one bullet, no speculation):

• docs/patterns/structured-dispatch.md — when to let the LLM call tools vs. when to make it classify only and dispatch deterministically.

---

Examples

Two canonical demo companies, each with one production-grade agent:

• examples/acme-creators/ — agency-managed personal-brand creator (Instagram setter)
• examples/stark-distributors/ — B2B distributor (stock/inventory assistant)

Read these before building. They are the source of truth for what good looks like.

---

Why agentstack and not [thing]

	Foco	What it doesn't do
Mastra	Agent runtime primitives	Single-tenant. No build workflow
LangGraph	Graph orchestration	Lower-level. No tenancy model
CrewAI	Multi-agent within one project	Not multi-client
agentstack	Multi-client agent workflow + framework	(the wedge)

agentstack is for agencies, consultancies, internal AI teams, and anyone who builds agents for more than one stakeholder.

---

Contributing

PRs welcome. See CONTRIBUTING.md. Iron Laws apply to contributions.

---

License

MIT — see LICENSE.