La agencia detrás de tu agencia.
Control-room SaaS for boutique marketing agencies. A 6-agent AI pipeline turns a brand brief into a full month of approved, ready-to-publish content — with humans approving every step.
Agentcy orchestrates six specialized AI agents in a strict sequential pipeline. Each agent produces output that the agency team reviews, edits, and approves before the next one runs. Nothing is auto-approved.
| Step | Agent | Output |
|---|---|---|
| 1 | Brand Intake | Brand Profile — voice, audience, pillars, visual kit |
| 2 | Brief | Campaign brief with objectives and KPIs |
| 3 | Strategy | Full content strategy with channel mix |
| 4 | Calendar | Editorial calendar with post schedule by channel |
| 5 | Copywriter | Post-by-post copy: hook, body, CTA, hashtags, video script |
| 6 | Reporte | Campaign performance report |
src/
├── agents/
│ ├── provider-registry.ts # Single LLM abstraction (Anthropic / OpenAI / Google)
│ ├── brand-intake.agent.ts
│ ├── brief.agent.ts
│ ├── calendar.agent.ts
│ ├── copywriter.agent.ts
│ ├── strategy.agent.ts
│ ├── report.agent.ts
│ └── prompts/ # System prompts as TS modules
├── app/
│ ├── (auth)/login/ # Two-column login, dark left panel
│ ├── (dashboard)/
│ │ ├── clients/[id]/ # Client detail + pipeline tracker
│ │ ├── clients/[id]/strategy/
│ │ ├── clients/[id]/calendar/
│ │ ├── clients/[id]/copywriter/
│ │ └── training/ # Fine-tuning dashboard
│ └── api/
│ ├── agents/ # Streaming agent endpoints
│ ├── strategies/
│ ├── editorial-calendars/
│ ├── copywriting-projects/
│ ├── trajectories/ # Data flywheel
│ ├── datasets/
│ └── train/
├── components/
│ ├── agents/approval-panel.tsx # Shared approval UI (all 6 steps)
│ ├── brand-intake/
│ ├── calendar/
│ ├── copywriter/
│ ├── strategy/
│ └── training/
└── types/ # Zod-validated shared types
Provider Registry — src/agents/provider-registry.ts is the only file that imports from @ai-sdk/anthropic, @ai-sdk/openai, or @ai-sdk/google. Every agent receives an AgentProvider instance and never knows which LLM is behind it. This makes model-swapping a one-line config change.
Human-in-the-loop — No agent output is ever persisted with status: 'approved' automatically. The pipeline gate is strict: agent generates → user reviews/edits → user approves → next step unlocks.
Streaming — All agents stream via Vercel AI SDK (streamText on the API route, useCompletion on the client). Users see output appear token-by-token.
Multi-tenancy — Every Supabase query filters by agency_id explicitly. RLS is enabled on all tables as a second line of defense, but it's never the only guard.
Data flywheel — Every agent interaction is captured as a trajectory (input, output, segments, tokens). Users rate outputs; approved edits become training signal. The training dashboard generates fine-tuning datasets from high-quality trajectories.
| Concern | Choice |
|---|---|
| Framework | Next.js 15 (App Router) |
| Language | TypeScript strict |
| Database | Supabase (Postgres + RLS + Auth) |
| AI SDK | Vercel AI SDK — streamText / generateText |
| LLMs | Anthropic Claude (primary), OpenAI, Google |
| Validation | Zod |
| Styling | Tailwind CSS v4 + shadcn/ui |
| Fonts | Plus Jakarta Sans · DM Sans · JetBrains Mono |
| CLI agent | Hermes (HERMES_CLI_PATH) |
- Node.js >= 20.9.0
- A Supabase project
- At least one LLM API key (Anthropic recommended)
cp .env.local.example .env.localNEXT_PUBLIC_SUPABASE_URL=your-supabase-url
NEXT_PUBLIC_SUPABASE_ANON_KEY=your-anon-key
SUPABASE_SERVICE_ROLE_KEY=your-service-role-key
ANTHROPIC_API_KEY=your-key
OPENAI_API_KEY=your-key # optional
GOOGLE_GENERATIVE_AI_API_KEY= # optional
HERMES_CLI_PATH=/path/to/hermes # optional, for CLI agentnpm install
npm run devOpen http://localhost:3000.
npx supabase db pushArchitecture rules, coding conventions, and the human-in-the-loop contract are in CLAUDE.md. Read it before writing any code — it applies to humans and AI agents alike.
Design tokens, component rules, and copy style are in DESIGN.md.
If you're using an AI coding assistant (Claude Code, Hermes, or other), see AGENTS.md — it mandates reading CLAUDE.md before starting any session.