diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 0000000..8b60405
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,69 @@
+# EXIF Test — Agent Guide
+
+EXIF + OCR test backend that validates GPS data survives the upload pipeline. Photo upload with metadata preservation, OCR machine ID reading, session management, and GPS proximity search.
+
+## Stack
+
+- **Python** 3.11+, **FastAPI**, **uvicorn**
+- **Pillow**, **pillow-heif** — image processing (JPEG, HEIC/HEIF)
+- **pytesseract** — Tesseract OCR
+- **Optional LLM OCR** — OpenCode Go / Google Gemini
+- **Frontend:** vanilla JS SPA, IndexedDB offline queue
+- **Service worker** (`static/sw.js`) — offline photo queue
+
+## Project Structure
+
+```
+server.py                 # ~1.5K lines — ALL FastAPI routes in one file
+new_endpoints.py          # Placeholder for future endpoints
+FEATURE_PLAN.md           # ~629 lines — comprehensive roadmap (✅ done / 💡 proposed)
+requirements.txt
+test_e2e_smoke.sh         # End-to-end smoke test
+
+static/
+  index.html              # SPA frontend (~79KB)
+  sw.js                   # Service worker for offline photo queue
+```
+
+## How to Run
+
+```bash
+# Dev (with reload)
+uvicorn server:app --reload --host 0.0.0.0 --port 8903
+
+# Production
+python server.py
+# or
+uvicorn server:app --host 0.0.0.0 --port 8903
+```
+
+## Key Architecture
+
+- **Single-file backend** — `server.py` has all routes
+- Hotels dotenv from `~/.hermes/.env`
+- Uses `photos.db` for photo/session metadata
+- Loads machine data from `canteen-asset-tracker/assets.db`
+- Supports Tesseract and optional LLM-based OCR (OpenCode Go, Google Gemini)
+- Frontend has IndexedDB offline queue for uploading photos without connection
+
+## Features
+
+- Photo upload with EXIF metadata extraction
+- OCR machine ID reading (ConnectID sticker)
+- Session-based grouping
+- GPS proximity search
+- Export: CSV, KML, clipboard
+- Bulk assign to machines
+
+## Production
+
+- **Port:** 8903
+- **No systemd**
+- **Dev only** — `--reload` in dev scripts
+
+## Pitfalls
+
+- **No requirements.txt committed** — dependencies inferred from imports
+- **Single-file backend** — server.py is ~1.5K lines
+- LLM-based OCR requires OpenCode Go or Gemini configured in environment
+- HEIC support requires `pillow-heif` + system libheif