OCI MCP – Agente de Consultas Imobiliárias OC
odanree/oci-mcp
Consultas imobiliárias OC em linguagem natural via Claude
Ver no GitHubServidor MCP que expõe um agente LangGraph de inteligência imobiliária como ferramenta para Claude Desktop e Claude Code. Permite fazer perguntas em linguagem natural sobre parcelas imobiliárias do Condado de Orange (endereços, APNs, proprietários, cadeias de títulos) e receber respostas fundamentadas com citações e rastreamento via Langfuse.
Instalação
npx convez-ai odanree/oci-mcp1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Pesquisar proprietários e históricos de imóveis no Condado de Orange
- Rastrear cadeias de títulos e portfólios de LLCs imobiliários
- Integrar inteligência imobiliária como ferramenta em fluxos Claude Code/Desktop
- Comparar propriedades e proprietários com citações verificáveis
Veredito

MCP bem arquitetado que encapsula um agente LangGraph robusto. Aplicável principalmente para profissionais imobiliários ou pesquisadores do Condado de Orange; a documentação é clara e a integração com Claude Code/Desktop é direta. Excelente exemplo de composição agentica e separação upstream-MCP.
README
A tradução para o português deste README ainda está na fila.
oci-mcp
MCP server that exposes the OC Real Estate Intel agent as a tool to Claude Desktop, Claude Code, and any other MCP client. Lets you ask natural-language questions about Orange County parcels — addresses, APNs, owners, title chains, LLC portfolios — and get back grounded answers with APN citations and a Langfuse trace ID.
MCP · FastMCP · Claude Desktop · Claude Code · Anthropic · LangGraph (upstream) · Pydantic v2 · pydantic-settings · httpx · respx · pytest-asyncio
| Tool | What it does |
|---|---|
oc_parcel_query | Natural-language query → answer + intent + citations + trace_id. The upstream agent does intent routing internally (lookup / summarize / compare / title_chain / portfolio); the MCP caller doesn't pick a mode. |
oc_submit_feedback | Attach +1 (thumbs-up) or -1 (thumbs-down) to a trace. Lands in Langfuse as a user_feedback score on the trace, so the agent's operator can filter "show me every query users marked bad." |
The upstream agent is documented at oc-realestate-intel and runs live at oci.danhle.net.
Use it from Claude Code
git clone https://github.com/odanree/oci-mcp
cd oci-mcp
python -m venv .venv && .venv/Scripts/activate
pip install -e ".[dev]"
cp .env.example .env
# Edit .env: set OCI_API_URL=https://your-deployment-url
Register the server (user scope = available in every Claude Code project):
claude mcp add oci -s user -- "C:\Users\Danh\Documents\Projects\oci-mcp\.venv\Scripts\python.exe" -m server.main
Verify:
claude mcp list
You should see oci: ... - ✓ Connected. Restart Claude Code, then in any chat:
"oci: who owns 461-211-62?"
"oci: what does FLORES FAMILY TR own?"
"oci: trace the title chain on 73 Bridgeport Rd Irvine"
Claude calls oc_parcel_query, the agent runs end-to-end (router → retrieval → summarize), and you see the answer cited by APN.
Use it from Claude Desktop
Same config, different file. Drop this into %APPDATA%\Claude\claude_desktop_config.json (Windows) or ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):
{
"mcpServers": {
"oci": {
"command": "C:\\Users\\Danh\\Documents\\Projects\\oci-mcp\\.venv\\Scripts\\python.exe",
"args": ["-m", "server.main"],
"env": {
"OCI_API_URL": "https://your-deployment-url"
}
}
}
}
Quit Claude Desktop fully, reopen, and the 🔧 icon should list the two tools.
Why this exists
Three reasons one of my MCP servers wraps my own agent system:
-
It turns a destination into a capability. Without this server, oc-realestate-intel is "a thing users go to." With it, the agent becomes a tool other agents can call. Different architecture, different composability.
-
It closes a real loop. I built an agent → deployed it to my infra → wrote an MCP that exposes it → other agents (or me, in Claude Code) call it. Every layer is mine, every layer is public, every layer is traced.
-
It's exactly what interviewers ask about. "How do you compose agentic systems?" gets a concrete answer: this server. Plus the upstream-vs-MCP separation maps directly to the read-vs-write split that infra-mcp uses, so the design story is consistent across the portfolio.
Tests
pytest
# 11 tests covering success path, all four error_kind cases (config / network /
# http / validation), citation normalization, bearer-token wiring, and feedback
# score validation. All HTTP mocked with respx — no network.
Project layout
server/
main.py FastMCP entrypoint, registers oc_parcel_query + oc_submit_feedback
config.py pydantic-settings, .env loaded via absolute path
tools.py Both API wrappers + Pydantic response models
tests/
test_tools.py respx-mocked tests for both tools
License
MIT