
mecenai
io.github.MonoVarTon/mecenai
Artist catalogue for AI agents: artworks, artist-approved prices, viewing requests.
Documentation
Mecenai
An AI sales concierge for artists and galleries. Mecenai answers collectors 24/7 in their own language, knows every artwork in the catalogue, books viewings and hands the conversation to a human at the right moment.
Mecenai — from mecenate, a patron of the arts, + AI. It connects artists with the collectors who will support them.
▶ Live demo: mecenai.onrender.com · MCP endpoint: https://mecenai.onrender.com/mcp/ (the free instance sleeps when idle — the first request may take up to a minute)
Status: v0.4 (November 2026) — concierge on a real LLM (Groq), model evaluation, an MCP server for other AI agents, and a public demo. This README describes where the product is going; each capability is marked with its current state.
Why
An artist's or gallery's website is often a beautiful catalogue that cannot answer a question. Collectors write at night, in other languages, and wait days for a reply. Mecenai turns the website into a conversation that leads to a viewing and a sale — in the artist's own voice, under the artist's own rules.
What it will do
| Capability | How | State |
|---|---|---|
| Knows every artwork: story, technique, size, series, intent | retrieval over the catalogue (RAG, vector search) | planned |
| Speaks in the artist's voice, in the visitor's language (English, Arabic, Russian, …) | LLM with artist-approved instructions | v0.2 — gpt-oss-120b via Groq |
| Quotes prices only by rules the artist approved — never negotiates or invents | tool calls with hard rules + tests | v0.1 — scripted model |
| Understands who is asking (collector, curator, press) and guides to a viewing | agent with state | planned |
| Hands off to a human in Telegram / WhatsApp | webhooks, n8n | planned |
| Exposes the catalogue to other AI agents | MCP server | v0.3 — tools, resource, prompt |
| Optional certificate of authenticity on TON | TON, IPFS | planned |
| Insights for the artist: what collectors ask about | analytics | planned |
| Measured quality: no invented prices or facts | automated evaluations | started — model comparison |
| One engine, many clients — each with its own catalogue and voice | multi-tenant service, Docker, cloud | planned |
Run it (v0.1)
python -m venv .venv && source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload
curl localhost:8000/health
curl -X POST localhost:8000/chat -H "content-type: application/json" \
-d '{"message": "What do you have from the Quiet Harbors series?"}'
v0.1 runs offline on a scripted model that speaks the Messages API format, so it needs no API key. The catalogue belongs to a fictional artist.
app/
main.py FastAPI: GET /health, POST /chat
concierge.py agent loop: model asks for a tool -> code runs it -> model continues
tools.py find_artworks, get_price_rule (artist-approved wording only)
models.py scripted model for tests + any OpenAI-compatible LLM (Groq)
catalog.json fictional demo catalogue
mcp_server.py MCP server: tools, resource, prompt
scripts/ model comparison
docs/evals/ evaluation results
tests/ API, price rules, hand-off, model adapter, MCP server
MCP server
Any MCP client (Claude Desktop, IDEs, other agents) can use the catalogue:
| Kind | Name | What |
|---|---|---|
| tool | find_artworks | search by series or title, facts only |
| tool | get_price_rule | the only way to get a price — artist-approved wording |
| tool | request_viewing | ask the artist for a viewing (validated, the artist replies personally) |
| resource | catalog://artist | the whole catalogue without prices |
| prompt | recommend_artwork | pick works for a collector's wishes |
Remote (any client that supports Streamable HTTP): https://mecenai.onrender.com/mcp/
Local, Claude Desktop (claude_desktop_config.json):
{"mcpServers": {"mecenai": {"command": "python", "args": ["/path/to/mecenai/mcp_server.py"]}}}
Real model: copy .env.example to .env, set MECENAI_MODEL=groq and GROQ_API_KEY.
Tests: pytest -q — they also run on every push (GitHub Actions).
Roadmap
| Version | What |
|---|---|
| v0.1 | service skeleton, tool calls, price rules, tests — done |
| v0.2 | real LLM backend behind the same interface; model comparison — done |
| v0.3 | MCP server: the catalogue for other AI agents — done |
| v0.4 | public demo and MCP over HTTP (Render), abuse limits — done |
| v0.5 | retrieval over artwork stories (RAG) |
| v0.6 | collector profiles and hand-off to Telegram |
Principles
- No invented facts or prices. If the answer is not in the catalogue or the rules, Mecenai says it will check with the artist and hands off.
- Honest about being an AI when asked.
- Collectors' data is minimal and private — privacy policy (including GDPR) from day one.
- Every number in this README comes from a measurement.
Open and closed
The engine, the MCP server, tests and evaluation results are open. Clients' catalogues, prices and conversations are never published; demos use a fictional artist.
Built by
Alexey — architecture, backend, automation, TON — at 369 Aether, a studio building websites for artists and galleries worldwide.
No installation packages available.