DivvyAI archives Divvy's public station feeds byte for byte, models them in BigQuery with dbt, forecasts the next hour and answers your questions through read-only MCP tools.
captures a day, one every fifteen minutes
stations in every capture
records loaded into BigQuery per run
duplicate rows after a replay
Product
Four layers, each small enough to understand and each tested on its own.
01 · Archive
Each run fetches station_information and station_status, then writes both responses and a manifest to one ZIP in Cloud Storage before anything is parsed. Replays reuse the original load job, so nothing is counted twice.
02 · Model
Staging views type the raw JSON. A snapshot fact joins status to station details from the same capture. Marts roll the snapshots up to hourly availability and measure the time between captures.
03 · Ask
A read-only MCP server gives Claude, or any MCP client, four tools: availability, capture gaps, station rankings and forecasts. Every query is capped at 256 MiB billed and nothing is ever written.
04 · Forecast
A gradient-boosted model predicts bikes and docks one to three hours ahead, with intervals. It trains only on data that was available at prediction time and beats both baselines.
Pipeline
Select a step to see how it is configured. Packets show a capture moving through the system.
Dashboard
Hourly averages from the warehouse, straight from mart_station_hourly_availability. Hover the chart to read each hour.
Network
Every dot is a station sized by capacity. Scrub through a weekday and watch bikes flow into the Loop for the morning commute, then back out at night.
Forecasts
One to three hours ahead for every station, with an interval and a probability of running out.
Model card
| Model, +1 hour | MAE (bikes) | Brier, empty |
|---|---|---|
| Persistence (last value) | 3.92 | 0.112 |
| Same hour last week | 3.11 | 0.094 |
| DivvyAI gradient boosting | 2.18 | 0.067 |
Monday 8:00 AM watchlist
MCP
Connect the DivvyAI MCP server to Claude and ask in plain language. Each answer comes from a typed, read-only tool call you can inspect.
get_station_availabilityread-onlyOne station's hourly bike and dock averages with sample counts. Up to seven days, 168 hours at most.
station_id, start, end
find_capture_gapsread-onlyPer-feed intervals between observed captures since yesterday, largest first.
min_minutes = 30, limit = 20
rank_stationsread-onlyRank stations by how often they ran empty or full at one Chicago hour, by weekday or weekend and area.
metric, hour, limit, start_date, end_date, day_type, area
forecast_availabilityread-onlyNext one to three hours of bikes and docks, with an 80% interval and the chance of running out.
station_id, horizon_hours = 3
Add it to Claude Code
claude mcp add divvyai -- uv run --directory ai python server.py
Reliability
Every scheduled slot for the last seven days. One square per capture, one row per UTC day.
Warns when the newest receipt is 30 minutes old and fails at 60. Runs hourly at minute 10.
Keys present and unique per capture and station; bike and dock counts present and never negative.
Any failed collector or checker run opens an incident in Cloud Monitoring and emails the on-call inbox.
Replays verify the saved archive and reuse the original load job instead of appending rows again.
Incident · October 18, 2026 · Resolved
The 16:45 UTC launch failed before the container started: Artifact Registry returned HTTP 401 during image import. Both feeds show a 29.98-minute gap from 16:31 to 17:01. The 17:00 run used the same image and loaded 4,148 records.
Developers
Python 3.12 with uv, Ruff, mypy and pytest. The default suite uses synthetic fixtures, so it never calls a live feed or the cloud.
uv sync --lockedLive availability, tomorrow's crunch points and every capture since launch.
