sodabar 🥤

An MCP server that puts open data on tap — four tools that let any LLM client search, inspect, and query 30,000+ civic datasets, with guardrails designed for a model on the other end.

Live demo: usmar-sodabar.static.hf.space — the same console, answered in your browser straight from the live NYC Open Data API.

• The problem — and why it matters
• What an agent session looks like
• The four tools
• The playground
• Key design decisions
• Limitations
• Reproduce it
• Project layout

The problem — and why it matters

Socrata powers the open-data portals of NYC, Chicago, Seattle, and hundreds of other governments — tens of thousands of live, queryable datasets. But an LLM can't use any of it directly: it doesn't know which datasets exist, what their columns are called, or how to write a SoQL query against them. And if you naively hand a model a raw HTTP tool, it will page 500,000 rows into its own context window or paste an HTML error page into its reasoning.

sodabar is a Model Context Protocol server that closes that gap. It exposes the catalog → schema → query workflow as four typed tools, with the sharp edges filed off server-side: row caps a tool call cannot exceed, dataset-id validation that fails before a request leaves the machine, and upstream errors rewritten into messages a model can act on ("check the dataset id and domain") rather than tracebacks it will hallucinate around.

Point Claude Desktop, Claude Code, or any MCP client at it:

{
  "mcpServers": {
    "sodabar": {
      "command": "/path/to/sodabar/.venv/bin/python",
      "args": ["-m", "sodabar.server"]
    }
  }
}

What an agent session looks like

docs/agent-demo.md is a committed transcript of Gemini driving the server through the real MCP stdio transport. Given only the four tools and the question "Which NYC borough has logged the most 'Rodent' 311 complaints so far in 2026?", the model planned three calls on its own:

1. search_datasets("311 Complaints") → found erm2-nwe9
2. get_schema("erm2-nwe9") → learned complaint_type, created_date, borough
3. query_dataset(select="borough, count(*)", where="complaint_type = 'Rodent' AND created_date BETWEEN …", group="borough", order="… DESC", limit=3)

and answered: Brooklyn 5,521 · Manhattan 3,528 · Queens 3,079. No SoQL was written by a human at any point.

docs/demo-transcript.md is the scripted equivalent — every tool exercised over a real stdio subprocess session, regenerated with make demo.

The four tools

Tool	What it answers
search_datasets(query, domain, limit)	"What datasets exist about X?" — full-text catalog search
get_schema(dataset_id, domain)	"What columns can I query, and what are their types?"
query_dataset(dataset_id, select, where, group, order, limit, offset, domain)	SQL-shaped aggregation and filtering via SoQL
profile_column(dataset_id, column, top)	"What values does this column take?" — vocabulary before where clauses

domain defaults to data.cityofnewyork.us but accepts any Socrata portal (data.seattle.gov, data.cityofchicago.org, …), so one server covers hundreds of cities.

The playground

make serve starts a FastAPI app whose REST routes mirror the MCP tools one-to-one — the console shows the exact tools/call envelope and the exact result an LLM client would see:

Guardrails are part of the demo. A malformed tool call gets a readable, actionable error — not a traceback:

The live static deployment serves the identical HTML with a 4 KB fetch shim that answers the /api/* routes in-browser, straight from the Socrata APIs (which send Access-Control-Allow-Origin: *) — a zero-backend demo of a backend project.

Key design decisions

• Guardrails live server-side, not in the prompt. $limit is clamped to 1,000 rows no matter what the model asks for; dataset ids must match Socrata's xxxx-xxxx form (which also blocks path traversal through the resource URL); domains must be bare hostnames. A prompt can be ignored — a clamp cannot.
• Errors are written for the model that reads them. A 404 becomes "not found — check the dataset id and domain"; a SoQL 400 surfaces Socrata's own message with "check your SoQL syntax". The retry policy distinguishes transient failures (429/5xx: three attempts with backoff) from semantic ones (400/404: fail immediately).
• One client, three consumers. The MCP server, the FastAPI playground, and the demo scripts share one SocrataClient, so timeout, retry, and error behavior can't drift between what's tested and what's deployed.
• Tool descriptions teach the workflow. The server's instructions and each tool's docstring steer a model toward search → schema → query and toward aggregating with group instead of paging raw rows — the difference between a 6-row answer and a 6,000-row context spill.
• Tests mock the transport, not the code. All 54 tests run against httpx.MockTransport — CI needs no network and finishes in under a second, while retry logic, error mapping, and the FastAPI lifespan wiring are exercised for real.

Limitations

• SoQL clauses are passed through to Socrata after shape checks, not parsed — a syntactically valid but expensive query (e.g. group on a high-cardinality column) is bounded by the row cap and Socrata's own timeouts, nothing stricter.
• Catalog search relies on Socrata's relevance ranking, which favors title matches; an agent may need two searches with different phrasings.
• Anonymous (keyless) Socrata access is throttled upstream; sustained heavy use would need an app token, which the client doesn't currently send.
• The committed transcripts hit the live API, so re-running make demo will show current counts, not the committed ones.

Reproduce it

git clone https://github.com/UsmarHaider/sodabar && cd sodabar
make venv        # python3 -m venv + editable install
make test        # 54 tests, no network needed
make demo        # scripted MCP stdio session → docs/demo-transcript.md (live API)
make serve       # playground at http://127.0.0.1:8012

cp .env.example .env   # then fill in GEMINI_API_KEY to run:
make agent-demo  # Gemini plans the tool calls → docs/agent-demo.md

Project layout

sodabar/
├── sodabar/
│   ├── soql.py            # query validation: 4x4 ids, domain shape, row caps
│   ├── client.py          # shared Socrata HTTP client: retries, error translation
│   ├── server.py          # the MCP server (4 tools, stdio transport)
│   ├── service.py         # FastAPI playground mirroring the tools over REST
│   └── web/index.html     # self-contained console UI (no build step, no CDN)
├── scripts/
│   ├── demo_session.py    # scripted MCP client session → docs/demo-transcript.md
│   ├── agent_demo.py      # Gemini function-calling over the MCP session
│   ├── screenshot.sh      # headless-Chrome captures of the console
│   └── deploy_space.py    # builds + publishes the static HF Space demo
├── tests/                 # 54 tests, all offline (httpx.MockTransport)
└── docs/                  # committed transcripts + UI screenshots