🇲🇾 Malaysia Data MCP Server

Production-grade Model Context Protocol server giving any AI assistant live access to Bank Negara Malaysia and data.gov.my datasets — over MCP, REST, and as a LangGraph agent tool.

Disclaimer: Unofficial community project. Not affiliated with or endorsed by Bank Negara Malaysia, DOSM, or the Government of Malaysia. Uses their free, public APIs.

---

What this is

Most AI assistants don't have live access to Malaysian financial data. Ask Claude or ChatGPT for today's OPR, this week's RON95 price, or whether a company is on Bank Negara's unauthorised list — you'll get a guess from training cutoffs or a refusal.

This server fixes that for any MCP-compatible client (Claude Desktop, Cursor, VS Code Continue, MCP-aware LangGraph), plus a REST API for everything else (n8n, Zapier, Streamlit, plain LangChain agents).

You: "What's Malaysia's current OPR and how does it affect Islamic home financing?"
LLM: [calls get_overnight_policy_rate via MCP]
     [calls get_islamic_interbank_rate via MCP]
     "Malaysia's OPR is currently 2.75% (effective 2026-05-07).
      Islamic interbank overnight rates are tracking at..."

Three transports, same 15 tools

Client	Transport	Use case
Claude Desktop, Cursor, VS Code	MCP stdio	Local agentic workflows
LangGraph, MCP-aware LangChain	MCP via langchain-mcp-adapters	Programmatic agents (see examples/langgraph_agent.py)
n8n, Zapier, dashboards, curl	REST on :8000	Anything HTTP

All three are backed by the same application-layer tool functions — single source of truth, no duplication. See ADR-0005.

Tools (15)

Bank Negara Malaysia

Tool	Returns
get_exchange_rates	Buying/selling/middle rates for major currencies
get_overnight_policy_rate	Current OPR + last change
get_base_rates	Base rates and BLR/BFR by bank
get_interbank_rates	Conventional interbank rates by tenure
get_islamic_interbank_rate	IIMM rates by tenure
get_kijang_emas_price	Live gold prices + per-gram derivation
check_consumer_alert	Whether an entity is on BNM's unauthorised list
get_usd_myr_reference_rate	KL USD/MYR reference rate (3:30pm daily)

data.gov.my (DOSM)

Tool	Returns
get_fuel_prices	Weekly RON95/RON97/Diesel + week-on-week change
get_cpi_inflation	Latest headline CPI + y-o-y inflation
get_gdp_data	Latest quarterly real GDP growth
get_population_stats	Population by state
get_household_income	Median + mean income by state

Derived & Composite

Tool	Returns
get_zakat_nisab_threshold	85g gold value in MYR (factual; not zakat advice — see ADR-0004)
get_malaysia_economic_snapshot	Composite: OPR + USD/MYR + gold + fuel + inflation in one call

Plus 3 MCP prompts (economic_briefing, scam_check, currency_planner) and 4 MCP resources for URI-addressable cached data.

---

Why this matters technically

Most MCP servers are 100-line scripts wrapping a few API calls. This one demonstrates the production patterns AI engineering teams ship in 2026:

• Hexagonal architecture — domain/ → infrastructure/ → application/ → presentation/. Tools are pure async functions, transports are thin wrappers, business logic is independently testable.
• Resilient HTTP — httpx async + tenacity retries with exponential backoff + jitter + aiolimiter token-bucket rate limiting + purgatory circuit breaker. Every upstream call is fault-tolerant by construction.
• Two-tier cache — L1 in-memory + optional L2 Redis with stampede protection via in-flight asyncio.Future sharing. Tunable per-tool TTLs.
• Observability triad — structlog JSON logs with correlation IDs + OpenTelemetry traces (auto-instrumented httpx) + Prometheus metrics on tool calls, latency, cache hit rate, circuit state.
• Pydantic v2 throughout — every API boundary validated, FastMCP auto-generates JSON Schema for tool I/O from the same models that validate REST.
• Strict typing — mypy --strict clean. No Any leaks.
• Real test discipline — 16 passing tests across unit (mocked HTTP via respx), property-based (hypothesis), integration (live APIs, optional), contract (daily upstream drift detection), plus an eval suite with 30 questions and automated scoring.
• 12-factor config — every setting from env vars via pydantic-settings. Single .env.example.
• Production deploy — multi-stage Dockerfile (~150MB), docker-compose stack with Redis + Prometheus + Grafana with a pre-built dashboard.
• CI — GitHub Actions matrix (3.10/3.11/3.12), ruff + mypy + tests + security scan + image build, plus a daily contract-check workflow that opens a GitHub issue if BNM/DOSM change their response shapes.

See docs/architecture.md for diagrams and docs/adr/ for the seven decisions that shaped the codebase.

---

Quickstart

1. Run the REST server (fastest demo)

git clone https://github.com/aliyaalias19/mcp-server-malaysia-data
cd mcp-server-malaysia-data
pip install -e .
python -m malaysia_data_mcp.presentation.http_server

Then:

curl http://localhost:8000/v1/opr | jq
curl http://localhost:8000/v1/snapshot | jq
open http://localhost:8000/docs   # Swagger UI

2. Run with the full observability stack

docker compose -f deploy/docker-compose.yml up

• Server: http://localhost:8000/docs
• Prometheus: http://localhost:9090
• Grafana: http://localhost:3000 (admin/admin) — dashboard pre-provisioned

3. Connect to Claude Desktop (MCP)

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%/Claude/claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "malaysia-data": {
      "command": "python",
      "args": ["-m", "malaysia_data_mcp.presentation.mcp_server"],
      "env": { "PYTHONPATH": "/ABSOLUTE/PATH/TO/mcp-server-malaysia-data/src" }
    }
  }
}

Restart Claude Desktop. You'll see malaysia-data in the tools panel.

4. Use as a LangGraph agent tool

pip install -e ".[agent]"
export ANTHROPIC_API_KEY=sk-ant-...
python examples/langgraph_agent.py

5. Run the eval suite

pip install -e ".[agent]"
export ANTHROPIC_API_KEY=sk-ant-...
python -m examples.eval_suite.runner

Reports per-category accuracy and tool-selection score on 30 hand-curated questions.

---

Testing

pip install -e ".[dev]"

pytest tests/unit tests/property            # fast, mocked        (default)
pytest tests/integration -m integration     # hits live BNM/DOSM
pytest tests/contract -m contract           # validates upstream API shape
pytest --cov                                 # coverage report

Test layout:

• Unit (tests/unit/) — every tool, mocked HTTP via respx
• Property (tests/property/) — Hypothesis invariants on parsing logic
• Integration (tests/integration/) — real BNM + data.gov.my calls
• Contract (tests/contract/) — daily-run upstream shape validation
• Smoke (tests/smoke/) — end-to-end MCP protocol smoke

---

Project layout

src/malaysia_data_mcp/
├── domain/               Pydantic models + error hierarchy
├── infrastructure/       HTTP, cache, observability, settings, upstream clients
├── application/          Tools, prompts, DI container
└── presentation/         FastMCP server + FastAPI server
tests/                    unit / property / integration / contract / smoke
examples/                 LangGraph agent + eval suite
deploy/                   Dockerfile, docker-compose, Prometheus, Grafana
docs/                     architecture diagrams + 7 ADRs + runbook

See ADR-0007 for the architectural rationale.

---

Configuration

All settings via env vars (MALAYSIA_DATA_*). Common ones:

Variable	Default	Purpose
MALAYSIA_DATA_ENVIRONMENT	dev	dev / staging / prod
MALAYSIA_DATA_LOG_JSON	true	structured vs human-readable logs
MALAYSIA_DATA_CACHE_REDIS_URL	unset	enables shared L2 cache
MALAYSIA_DATA_OTEL_ENABLED	false	OpenTelemetry tracing on/off
MALAYSIA_DATA_OTEL_ENDPOINT	unset	OTLP HTTP endpoint
MALAYSIA_DATA_HTTP_PORT	8000	REST bind port

Full list in src/malaysia_data_mcp/infrastructure/settings.py.

---

Related projects

• bnm-mcp — BNM-only MCP server (inspired this project)
• mcp-datagovmy — data.gov.my MCP server
• FastMCP — the framework underneath

This project unifies BNM + data.gov.my, adds production patterns (cache, circuit breaker, observability), exposes a parallel REST API, and ships with an eval suite.

---

License

MIT — see LICENSE.

---

Built by Aliya Alias — AI Engineer, Kuala Lumpur. MSc Artificial Intelligence, University of Malaya. Currently shipping production AI for Malaysian federal-government clients.