Singapore MCP: official Singapore data, inside your AI assistant

The problem it solves

Singapore publishes an unusual amount of high-quality public data. HDB resale transactions, ACRA company records, COE bidding results, MRT station and exit lists, live bus arrivals, PSI and rainfall readings, GDP and CPI series, MAS exchange rates, IRAS collections, visitor arrivals - almost all of it open, documented, and free.

Getting an AI assistant to use it properly is the hard part. Ask a general-purpose model about Toa Payoh resale prices and you get a plausible paragraph assembled from whatever it absorbed in training, with no dataset behind it and no date on it. The honest version of that answer requires finding the right dataset, filtering it to the right town, flat type and month, doing the arithmetic, and then saying out loud which agency published it and when.

Singapore MCP does exactly that, as a Model Context Protocol server: a small program your AI app runs, which gives it a set of tools it can call. You install it once and then ask questions in plain English.

Every answer shows its work

This is the design rule the whole project is built around. Wherever the upstream format allows it, a result keeps its source agency, dataset or table identifier, source URL, retrieval time, observation period, units, and freshness caveats. A stale or frozen series says so. A number whose units differ from the series next to it is not quietly averaged with it. A company-name match that could be ambiguous is reported as ambiguous.

That is what makes the output checkable rather than merely fluent - and it is the same instinct behind everything else we build: an answer a human can audit is worth more than an answer that merely sounds right.

What is in the box

• 291 read-only tools across property, rail, buses and roads, companies, weather and environment, education, the economy, and official financial reference data.
• Eight Agent Skills - packaged workflows that teach a client when to reach for which tool, for property, mobility, economy, business, civic and finance research, plus one for developing against the server itself.
• Seven Claude Code plugins: the complete suite, plus six focused profiles (property, mobility, business, economy, civic, finance) for sending fewer tool definitions to the model.
• Twelve packages: install the aggregate server for everything, or a focused executable when you only need one provider - data.gov.sg, OneMap, LTA, weather, or rail.
• Both transports: stdio for local clients, Streamable HTTP for the ones that prefer it. Plus a command-line tool for calling the same tools from a terminal.

Upstream sources are data.gov.sg (including the ACRA, HDB, URA, MOE, ECDA, IRAS, STB, EMA, SPF, MOH and NEA datasets in the curated catalog), OneMap from the Singapore Land Authority, LTA DataMall, and the SingStat Table Builder.

What it looks like in a chat

These two screenshots are real exchanges with the Olano WhatsApp agent, which answers Singapore questions with these tools. Note what the second reply does with a badly-behaved distribution: it gives the average, then the median, then the range, then explains the gap between them - newer Bidadari and Alkaff units against older stock - and flags that the current month is still filling up. The follow-up asks for a chart, and the agent builds one and sends it back into the chat.

Install it in your own AI client

Running it yourself is the everyday route: the server runs locally, has no Olano account, and sends no telemetry. You need Node.js 20 or newer for every route except the Claude Desktop extension, which carries its own runtime.

Claude Code

Use the plugin - it installs the MCP server and all eight Agent Skills, so answers come back better sourced. Two commands at the Claude Code prompt:

• /plugin marketplace add olano-ai/mcp-singapore
• /plugin install olano-singapore@olano

Choose User scope so it works in every project, then run /mcp to confirm singapore is connected. The first call takes 20-30 seconds while the package downloads; after that it is fast.

Claude Desktop app

Download olano-singapore.mcpb from the project's latest release and double-click it. Claude Desktop shows a review dialog; select Install, leave every optional key blank, and ask a question. No terminal, no restart.

Codex app and Codex CLI

They share one configuration. From the CLI: codex mcp add singapore -- npx -y @olano/mcp-singapore. In the app: Settings → MCP servers → Add server, name it singapore, choose STDIO, command npx, arguments -y @olano/mcp-singapore.

Anything else that speaks MCP

Point any client at npx -y @olano/mcp-singapore over stdio, or start it on HTTP with --transport http. The full instructions, the optional credentials, the profile list and the troubleshooting notes all live in the repository README.

Where this fits with Olano

Olano is a platform for AI agents that do real operational work - answering customers on WhatsApp and Telegram, keeping a CRM honest, running scheduled research - inside an isolated deployment with human approvals, audit history and hard spending caps. Those agents speak MCP, which is how the Singapore tools reach the demo bots above: the server is added to a deployment like any other MCP server, and the agent's approval rules and spend limits apply to it unchanged.

So this page is two things at once. It is a free tool for anyone in Singapore working with public data - use it, fork it, file issues. And it is a fair sample of how we build: one narrow job, done with the sourcing attached, so the person reading the answer can check it.

The boundaries we keep

• It is not a government product. Singapore MCP is an independent community project, not affiliated with, endorsed by, or an official product of the Singapore Government or any agency. "Official" describes the upstream data, never the software.
• It is read-only. All 291 tools query published APIs and datasets. There is no write path.
• The mortgage tools are educational. They surface official reference-rate statistics such as SORA and published housing-loan series, and calculate transparent illustrative scenarios. They are not live lender offers, credit decisions, or personal financial advice.
• Coverage tracks the upstream. When a dataset is stale or frozen, the tool says so rather than presenting an old number as current.