wiatrek-kevin-s/columbus

A governed GIS workbench for working with AI.
Your assistant reads, checks and maps your data under your organization's rules, and nothing changes or leaves your computer without you.
Download ·
Install on a Mac ·
Install on Windows ·
Technical reference
A chat in the Columbus app. The assistant read two layers, looked at them on a map to check that they line up, and answered. The map, the layers and their checks are beside the chat.
> [!WARNING]
> Test releases. Columbus is young software, built mostly by an AI model (see Disclaimer). Try it on data you can afford to have read, and have your organization review it before using it on production or sensitive data.
## Why Columbus
Your AI, your data, your rules. Use Columbus from Claude desktop or any other AI app that speaks MCP, or from its own chat with your Anthropic API key. Whatever the AI asks for, Columbus decides in its own code (not in a prompt) what it may read and do.
Read-only by default. Every connection is read-only until your organization's settings allow edits, and even then each change set is shown to you before anything is written.
Nothing leaves or changes without you. Exports, edits, scripts and downloads wait for your yes. Approvals are bound to the exact request, single-use, and expire.
Every result is checked. Row counts, CRS and units, geometry validity, row limits and filters come back with every result as checks: passed, to review, or failed. A failed check blocks exports until you decide.
It shows its work. Every output has a provenance record (inputs, steps, checks), and every call, including refusals, is a line in a hash-chained audit log on your computer.
Local and private. Columbus runs on your computer. Your data goes to the AI service you chose, and only the rows the rules let it see. Columbus itself sends nothing anywhere: no telemetry, no account.
## A tour
Chat : the assistant's steps fold into short lines with check badges; the map and table sit beside it.
Map first : one click puts the map in the middle and the chat on the side, for reviewing results.
Approvals : an export to a folder that isn't pre-approved waits for you, with exactly what it will write and where.
Home : what needs you, from every chat and AI app on this computer: requests, notes in draft, tools and workflows to review.
Workflows : a job your assistant proposed, in plain words and in full, approved once with Touch ID or Windows Hello.
Workflow runs : each run's report: steps, row counts, checks and what was exported.
Your own tools : a script your assistant packaged as a tool, read line by line before you install it.
Connections : your databases and files, and in plain words what the assistant may see in each.
Audit log : every call any assistant made through Columbus, and every decision, with a check of the chain.
Settings : the AI for chats (model and thinking level), the server and engine in use, and your organization's switches.
Every screenshot is taken from the app itself by a script (app/desktop/e2e/screenshots.py), on made-up data around Austin, Texas.
## The app, page by page
The Columbus app is one window with a menu down the side.