Scripts rank, the model judges: photo geolocation with receipts
An agent skill that finds where a photo was taken from pier spacing, shadows and ridge lines — and forces every conclusion to cite the command and file behind it.

What it does
Hand your agent a photo and one sentence; you get back camera coordinates with an error radius, the direction the camera was facing, a satellite evidence image, and a graded confidence. It’s a standard Agent Skill — SKILL.md plus plain Python scripts — so the same folder runs in Claude Code, Codex, Cursor, Gemini CLI, OpenCode and GitHub Copilot. The point: it works when the frame has nothing to read. No signs, no plates, no landmarks — OpenStreetMap geometry, elevation data, satellite tiles and street view carry the search on their own.
The interesting bit
The division of labor is the design: twenty single-purpose scripts search, score and sort, and the model only judges among the top few. The house rule is stricter still — a conclusion must name the command that ran in the session and the file it produced, because “population and fame are not evidence.” The showcase case: an EXIF-stripped photo of a rice paddy, a viaduct and a mountain gets narrowed from 27,335 OSM railway-bridge segments to 171 sites by horizon shape, then 14,372 rendered ridge lines are fitted to the photo, and finally 17 piers become 17 bearings whose spacing must come out even — one spot, ±2 m.
Key highlights
- Geometry does the work: pier spacing becomes a distance ruler, shadows become a bearing, a ridge line becomes a fingerprint matched against elevation data.
board.pyruns a candidate board with clues, likelihood ratios, exclusions and scan order; exclusions require read or computed evidence, while guesses can only lower a weight.- The toolbox spans EXIF, OCR (Apple Vision on macOS, RapidOCR elsewhere), reverse image search on Baidu and Yandex, sun/shadow math, and lookup tables for plate prefixes, area codes and driving side.
- Answers ship with an error radius and graded confidence, not bare coordinates.
Caveats
- The showcase run took about 72 minutes end to end, roughly half of it waiting on computation — this is patient work, not a lookup.
- The README is honest about its assumptions: the pier ruler used an assumed 32 m span, and the opening region guess (“South China, as a bet, not a proof”) came from vegetation clues.
- It’s a networked pipeline — OSM, elevation data, satellite tiles, street view, Baidu/Yandex — not something that runs cut off from the world.
Verdict
OSINT hobbyists, GeoGuessr degenerates, and anyone designing agent skills where the model must stay on a leash will get something out of this. If you want a fast geolocation API or a one-call library, this isn’t it — it’s a slow, deliberate investigator you install as a skill.
Frequently asked
- What is Oldcircle/geo-sleuth?
- An agent skill that finds where a photo was taken from pier spacing, shadows and ridge lines — and forces every conclusion to cite the command and file behind it.
- Is geo-sleuth open source?
- Yes — Oldcircle/geo-sleuth is open source, released under the MIT license.
- What language is geo-sleuth written in?
- Oldcircle/geo-sleuth is primarily written in Python.
- How popular is geo-sleuth?
- Oldcircle/geo-sleuth has 7.4k stars on GitHub.
- Where can I find geo-sleuth?
- Oldcircle/geo-sleuth is on GitHub at https://github.com/Oldcircle/geo-sleuth.