Start here

OpenDayton lets an AI assistant answer questions from the City of Dayton and

Montgomery County's own published data — crime, housing condition, lead water

lines, trash pickup, capital projects, and fifteen years of county property

records.

It's free, read-only, and takes about two minutes to connect.

1. Connect it (2 minutes)

Claude (web or desktop)

• 1. Settings → Connectors → Add custom connector
• 2. Name: OpenDayton — URL: https://opendayton.org/mcp
• 3. Save. In a new chat, click + → Connectors → switch OpenDayton on.

Claude Code

claude mcp add -s user --transport http opendayton https://opendayton.org/mcp

Anything else that speaks MCP — point it at https://opendayton.org/mcp

(Streamable HTTP, no key needed).

2. Ask it something

• Which neighborhoods had the most violent crime last year?
• Now show that as a rate per 1,000 residents.
• What day is trash pickup at 275 Linden Ave, and does it have a lead water line?
• How many demolition permits has the county recorded each year since 2015?
• Tell me everything you know about my address.
• What share of homes in Dayton are owner-occupied?

Try one about your own street. If the answer looks wrong, that's interesting —

tell us.

3. Run it with your own model instead (optional)

No API key, no cloud, nothing leaves your laptop. You need

Ollama.

**Do this before the workshop if you can — the model is a 5 GB download and

venue wifi will not enjoy twenty people doing it at once.**

ollama pull granite4.1:8b

Then either use a ready-made client:

uv tool install ollmcp
ollmcp -u https://opendayton.org/mcp -m granite4.1:8b

…or run our example, which prints every tool call so you can watch it think:

git clone https://github.com/codefordayton/opendayton && cd opendayton
uv sync
uv run python examples/local_model.py --model granite4.1:8b \
  "How many crimes were reported in Five Oaks in 2025?"

Tool calling, not writing quality, is what matters for this. Measured on an

M2 Max: granite4.1:8b (5.3 GB) answers in about 30 seconds and three tool

calls; a similarly sized model that isn't tuned for tools took four times

longer and made five.

4. Add something to it

The catalog is the product — sixteen datasets, each with its fields chosen by

a person and its traps written down. Adding one is editing a text file.

See CONTRIBUTING.md

for both paths: adding a dataset (YAML, no programming) or adding a tool (Python).

What it will not do

Owner names, the addresses of vacant buildings, exact arrest dates and ages,

and crime-victim details are deliberately not exposed — all of it is public

record, but public record and one-sentence-away are different things. The

reasoning is in DECISIONS.md.

Other pieces

• Parcel geocoder — https://geocoder-production-bb7e.up.railway.app/geocode?address=275+LINDEN

turns any Montgomery County address into a parcel ID and coordinates. Public,

read-only, free to use in your own project.

• Source — github.com/codefordayton/opendayton, MIT.

A Code for Dayton project.