CHE MCP — Argentina's First National MCP Ecosystem: 5-Stage Intelligent Gateway with WMA Online Learning

[Albano-schz]

Pre-submission Checklist

• This post follows the MCP community guidelines

What would you like to share?

What it is

CHE MCP is an intelligent gateway that connects AI agents with real-time data from Argentina — 80+ government datasets, finance APIs, weather, transit, football, and tax compliance — through a single MCP server.

Instead of installing 80 individual MCPs, you install one. The gateway understands natural language queries in Spanish and routes them automatically to the correct data source.

Architecture highlights

5-stage classification pipeline with graceful degradation:

1. Keyword matching — 3,000+ Spanish keywords across 182 classified domains
2. WMA weighted routing — Weighted Majority Algorithm with online learning (+0.1 reinforce, −0.1 penalize, bounded [0.1, 5.0]). Persisted to disk. Router improves with every query.
3. Semantic embeddings — 384-dim vectors (all-MiniLM-L6-v2 via @xenova/transformers) with Jaccard fallback for zero-dependency offline mode
4. Data Node search — DuckDB + NL-to-SQL on 748 Parquet datasets from datos.gob.ar (9.92× Zstd compression — 404 MB vs 3.92 GB CSV)
5. LLM fallback — optional external endpoint

Resilience patterns:

• 3-tier cache (in-memory LRU → persistent disk → live CKAN) with atomic file writes
• Per-dataset circuit breakers (3 failures → serve stale, 60s cooldown)
• Request collapsing (concurrent identical queries share a single upstream fetch)
• Predictive pre-fetch on top-10 hot datasets every 15 minutes

Protocol & compliance:

• Streamable HTTP MCP transport — ready for the July 2026 protocol update
• MCP SDK @modelcontextprotocol v1.29.0 with server.registerTool API
• JWT + API key auth with per-key rate limiting (100 req/min)
• OpenTelemetry distributed tracing
• 280+ tests (Vitest), Zod validation, TypeScript 5.4 + Node.js 24

Benchmark: 95.45% TFS accuracy on MCPAgentBench (66 diverse queries).

Status

Currently in active deployment. The Data Node with 748 Parquet datasets and DuckDB NL-to-SQL query engine is going live this week.

Full technical documentation is public: https://github.com/Albano-schz/che-mcp-docs

Built from Bahía Blanca, Buenos Aires, Argentina.

---

Happy to answer questions or get feedback from the MCP community.

Relevant Links

https://github.com/Albano-schz/che-mcp-docs