Decor Agent (Decora)

An MCP-first interior design agent. Decora furnishes a room from a real catalog, writes the job to a design project, and stops for human approval before anything is committed.

This is not a chat wrapper. The model does not invent SKUs or keep the spec in conversation history. The world lives on an MCP server; the Decor app is a host that discovers tools, reads resources, and lets you approve the cart.

Built against the Agentic AI Foundation MCP standard. Why each slice exists is in docs/aaif-learning.md.

What it does

You describe a room, budget, and constraints. Decora searches seeded inventory, updates project://{context_key}, and calls request_approval when the spec covers the room and the budget holds. You approve or reject in the UI. Draft items stay draft until you do.

Example jobs:

• "12x14 living room, $2000, keep grandma's credenza"
• "What paint from the catalog works with dark oak floors?"
• "Add a sofa under $1200 and pause for approval"

Old specialist prompts (style_advisor, room_planner, trend_spotter) are still in the tree. /api/chat does not use them.

Architecture

Host (FastAPI + web/)
  chat UI + project panel
  harness: observe → tools/list → model → tools/call → stop for approval
  MCP client
        │
        │  JSON-RPC (in-process Client, or stdio for Inspector)
        ▼
Server (mcp_servers/decor_design.py)  — no Claude
  tools:     search_catalog, update_project, request_approval
  resources: project://{context_key}, catalog://sku/{sku}
  prompts:   plan_room (user-invoked)
        │
        ▼
World: seeded catalog + in-memory project store

Primitive	Who controls it	In this repo
Tools	Model	Search inventory, mutate the project, ask to commit
Resources	Host / app	Read the project and a SKU without a tool call
Prompts	User	plan_room — room, budget, keep, avoid

The model cannot call approve. That is a host route on purpose.

Project layout

decor-agent/
├── mcp_servers/decor_design.py   # MCP environment (tools, resources, prompts)
├── app/
│   ├── catalog.py                # Seeded SKUs — if it is not here, it does not exist
│   ├── project.py                # Brief, rooms, spec list, budget, approval
│   ├── store.py                  # In-memory projects keyed by context_key
│   ├── harness.py                # Host loop: discover tools, call MCP, read project://
│   ├── prompts.py                # Decora system prompt (not plan_room)
│   ├── graph.py                  # Legacy specialist graph — unused by /api/chat
│   └── tools/                    # Legacy LLM “tools” — unused by /api/chat
├── server.py                     # FastAPI: /api/chat, /api/project, approve, reject
├── web/                          # Chat + project panel
├── test_catalog.py
├── test_store.py
├── test_mcp_catalog.py
├── test_mcp_project.py
├── test_harness.py
├── test_agent.py                 # Live LLM e2e (needs ANTHROPIC_API_KEY)
├── docs/aaif-learning.md
└── .env.example

Quickstart

python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env           # then add ANTHROPIC_API_KEY
python server.py               # http://localhost:8000

Open the chat UI and give Decora a room. The right-hand panel is project://, not a second transcript.

curl -X POST http://localhost:8000/api/chat \
  -H 'Content-Type: application/json' \
  -d '{"message": "Plan a 12x14 living room with a $2000 budget"}'

Host-only (the model cannot hit these):

curl 'http://localhost:8000/api/project?context_key=demo'
curl -X POST http://localhost:8000/api/project/approve \
  -H 'Content-Type: application/json' \
  -d '{"context_key": "demo", "kind": "spec"}'

Inspect the environment with no LLM

The MCP server is usable without the chat app:

python mcp_servers/decor_design.py

Or open MCP Inspector:

npx @modelcontextprotocol/inspector python mcp_servers/decor_design.py

Then tools/list, resources/read on project://demo or catalog://sku/ART-SOFA-721, and prompts/get plan_room.

Tests

No API key needed for the protocol and store tests:

python test_catalog.py test_store.py test_mcp_catalog.py test_mcp_project.py test_harness.py

Live routing against Claude:

LOG_LEVEL=WARNING python test_agent.py

Environment

Variable	Default	Purpose
ANTHROPIC_API_KEY	required for chat	Claude API key (not required for Inspector or MCP unit tests)
ANTHROPIC_WORKSPACE_ID	""	Required if the key is not scoped to one Anthropic workspace
LD_SDK_KEY	""	LaunchDarkly server SDK key (unused for the MCP loop)
LOG_LEVEL	INFO	structlog level
ENVIRONMENT	development	Console vs JSON logs

Unknown .env keys are ignored so leftover Temporal-era variables do not crash settings.

What's next (not in this tree)

AAIF build order after MCP + this host:

1. Point goose at decor-design (manual check — if goose cannot furnish a room, the server is not done)
2. AGENTS.md — stop conditions and consent, once the agent exists
3. A2A — e.g. a retailer agent for stock
4. agentgateway — only when there is more than one thing to front
5. LaunchDarkly — gate work that is already real

Temporal (API World durable-workflow demo) lives on temporal-api-world and durable-workflows, not on main.

Tech stack

Python 3.12 · MCP Python SDK (MCPServer) · Anthropic Claude Sonnet 5 · FastAPI · Pydantic · structlog