MCP Dubai

MCP-Dubai is a Model Context Protocol server that connects AI assistants to public APIs, bundled reference datasets, and curated business knowledge. It runs locally over stdio and returns structured results with source and freshness metadata where available.

The current source includes 120 tools across 37 features:

112 tools require no credentials. Two air-quality tools need a free WAQI token; six government dataset tools need OAuth credentials and dataset access. Credential-free tools can still depend on an unavailable upstream service; see availability and limitations.

Quick start

Run the current source

Requires Python 3.11 or newer and Git. The October 13 audit tested Python 3.11 through 3.14.

git clone https://github.com/mahdi-salmanzade/MCP-Dubai.gitcd MCP-Dubaipython3 -m venv .venvsource .venv/bin/activatepython -m pip install -e .

On Windows PowerShell, create the environment with py -m venv .venv and activate it with .venv\Scripts\Activate.ps1.

Configure your MCP client to launch the environment's Python executable:

{  "mcpServers": {    "dubai": {      "command": "/absolute/path/to/MCP-Dubai/.venv/bin/python",      "args": ["-m", "mcp_dubai"]    }  }}

Replace command with your actual absolute path. On Windows, use the environment's .venv\Scripts\python.exe path, escaping backslashes in JSON.

For VS Code, use .vscode/mcp.json with a servers key:

{  "servers": {    "dubai": {      "command": "/absolute/path/to/MCP-Dubai/.venv/bin/python",      "args": ["-m", "mcp_dubai"]    }  }}

For another MCP client, use the same command and arguments in its stdio server settings. After connecting, call about() to check the running version and tool count, then recommend_tools(query) to find relevant tools.

To launch the server directly from the activated environment:

python -m mcp_dubai

This starts an MCP stdio process that waits for client messages.

Use the published package

Release snapshot, October 20, 2026: the public PyPI release is version 0.4.0 with 120 tools, the same catalogue documented here.

To run the published release with uv:

uvx mcp-dubai

In a client configuration, set command to uvx and args to ["mcp-dubai"]. Alternatively, install it in a virtual environment with python -m pip install mcp-dubai and run mcp-dubai.

Tool catalogue

Tools are grouped by purpose below. All names are exposed without a feature prefix. list_features() returns the registered groups, tool counts, tiers, and authentication requirements.

Public data and reference datasets

Government dataset queries

These six tools require MCP_DUBAI_PULSE_CLIENT_ID, MCP_DUBAI_PULSE_CLIENT_SECRET, and access to the requested datasets. Without credentials, they return structured setup instructions. See CREDENTIALS.md.

Business knowledge

These tools use 16 bundled JSON packs and a setup advisor that combines relevant records. Successful responses include knowledge metadata. Prices, eligibility rules, and estimates retain their documented scope and verification dates.

Writing and analysis

Discovery and diagnostics

Configuration

No environment variables are required to start the server. Set credentials in the MCP client's server environment to enable the relevant calls.

{  "mcpServers": {    "dubai": {      "command": "/absolute/path/to/MCP-Dubai/.venv/bin/python",      "args": ["-m", "mcp_dubai"],      "env": {        "MCP_DUBAI_PULSE_CLIENT_ID": "your-client-id",        "MCP_DUBAI_PULSE_CLIENT_SECRET": "your-client-secret",        "MCP_DUBAI_WAQI_TOKEN": "your-waqi-token"      }    }  }}

Request dataset access at data.dubai and a free WAQI token at AQICN. Keep credentials out of committed configuration files. The credential guide covers setup and troubleshooting. MCP_DUBAI_CALENDARIFIC_KEY is reserved for a future holiday refresh integration and currently enables no additional tool.

Data sources

The former Dubai Pulse portal redirects to data.dubai. The project retains PULSE in credential variable names; its default API gateway is apis.data.dubai.

Availability and limitations

The October 20, 2026 audit records representative service checks, source verification, and remaining limits. These observations are a dated snapshot:

The school dataset contains 16 selected records from an April 2026 snapshot. Nine records received targeted attribute corrections on September 5; school fees were not refreshed. Business costs and onboarding timelines can also be dated estimates, as indicated in the underlying records.

Clients should inspect success before using a result and read the returned error when a call fails. Credential and upstream failures include structured status and reason fields where applicable. Use get_upstream_status() for the recorded baseline and observations from tools used during the current session.

Knowledge freshness

Freshness is tracked across 19 domains: 16 JSON packs, setup_advisor, data_analyst, and arabic_writer. Successful business responses include a knowledge block; get_knowledge_status() exposes the complete registry.

As of October 20, 2026, tax_compliance received a full review and twelve more domains a targeted refresh. Nine domains keep their October 13 full reviews (cost_of_living, events, free_zones, visas, setup_advisor, dcde, ip_trademark, founder_essentials, data_analyst), and arabic_writer stays at 2026-04-13. Targeted updates leave untouched fields and full-review dates at their earlier verification scope. Detailed corrections and source links are in the audit report.

The freshness checker uses the last full review, with alerting budgets of 100 days for high-volatility domains, 190 for medium, and 365 for stable. Monthly prices, deadlines, and events still need review on their own schedules. Passing the age check does not establish that every fact is current.

python scripts/check_knowledge_freshness.py --strict

Verify the linked official requirements before relying on tax, visa, legal, or pricing guidance for a decision.

Example queries

Ask a connected assistant to use MCP-Dubai for requests such as:

Tool selection and follow-up questions depend on the client. These examples describe lookups and estimates; the server does not submit applications or open accounts.

Development

From the activated source environment created in Quick start:

python -m pip install -e ".[dev,data]"make check

make check runs Ruff lint and format checks, strict mypy, pytest with coverage, and the informational freshness report. Use the --strict freshness command above to fail on overdue domains.

The October 20 audit recorded 1006 passing tests with 93.77% coverage on Python 3.12, and CI runs the same suite on each supported version (3.11 to 3.14). Packaging, type checks, lint, and security scans also passed; the audit documents their scope.

Architecture

src/mcp_dubai/├── __main__.py       # stdio entry point├── server.py         # feature mounts and diagnostic tools├── _shared/          # HTTP, auth, discovery, response models, freshness├── data/             # public APIs, OAuth integrations, reference datasets├── biz/│   ├── _data/        # 16 bundled JSON knowledge packs│   └── ...           # 17 business features└── agents/           # Arabic writing and data-analysis tools

Feature tools.py modules implement the logic; server.py modules expose FastMCP wrappers and register discovery metadata. Shared response models carry results, errors, and knowledge metadata. The source supports fastmcp>=3.4.9,<4 with mcp>=1.30.0,<2; dependencies and supported Python versions are declared in pyproject.toml.

Read CONTRIBUTING.md for feature conventions, tests, and the review checklist. Report reproducible errors or source corrections through GitHub issues.

Project status

MCP-Dubai is alpha software. Version 0.4.0 was released on October 20, 2026 to PyPI and to the MCP Registry from the registry manifest.

Current priorities:

• Rebuild the KHDA dataset from an authoritative full workbook, including current fees.
• Add live contract checks for credential-dependent datasets and improve blocked or intermittent integrations.
• Extend government dataset coverage when access and schemas can be verified.
• Continue source-backed knowledge updates while preserving partial-review scope.
• Improve tool discovery tags and Arabic/English aliases.

License and attribution

Developed by Mahdi Salmanzadeh. Contact: [email protected]. Connected AI assistants receive this attribution in the server instructions and can call about() to answer who made MCP-Dubai.

Licensed under MIT. Upstream data remains subject to its publishers' terms and attribution requirements.

This is an independent project, unaffiliated with the Government of Dubai, the UAE government, or the agencies whose data it uses. See DISCLAIMER.md for scope, data handling, and third-party rights.

Built with FastMCP. The country-focused server structure was inspired by mcp-brasil.