oldcircle/geo-sleuth is an agent skill that finds the shooting location and the direction the camera was facing from a single photo. Even when a photo has no sign, license plate, or other readable text, it narrows down candidates using OpenStreetMap geometry, elevation data, satellite tiles, and Street View imagery as evidence. It consists of SKILL.md and Python scripts, and it is not an MCP server.
The gap between bridge spans serves as a ruler, shadows give a bearing, and ridgelines are compared against terrain like fingerprints. Exploration, scoring, and ranking are handled by about twenty scripts, and the model only picks among the top candidates. Results include coordinates with an error radius, the camera direction, the evidence images, and a probability carrying a stage rating. Each conclusion is required to state the commands it ran and the files it generated during that session. The same folder works in Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and GitHub Copilot, and requires Python 3.10 or later. According to the README, no separate model is needed.
The README presents a case where the location was narrowed to within 2 meters from a photo showing fields, an elevated road, and a mountain behind a bridge. It does note, however, that other models may not reach the same result, and the repository contains no evaluation dataset confirming this accuracy and no reproducible measurement script. So the 2-meter case is a result recorded in the README, not a measurement that demonstrates general accuracy.