diff --git a/README.md b/README.md
index 89fd4a8..ffa8a59 100644
--- a/README.md
+++ b/README.md
@@ -6,6 +6,8 @@
 
 **Your self-hosted [StackChan](https://github.com/m5stack/StackChan) robot assistant — kid-safe by default, hackable by design, private by architecture.**
 
+> 🤖 **AI-assisted project.** Most of the code and nearly all of the docs in this repo were written by AI agents (primarily Claude Code) under my direction. I've been coding professionally for 15+ years; Dotty is one of a few side projects I'm using to learn the current generation of AI/LLM tooling first-hand — what the tools do well, where they break, and how to drive them. Feedback on the output very welcome.
+
 > ⚠️ **Heads up: this is not a stable project yet.** Dotty is buggy, frequently broken, and actively changing day-to-day. End-to-end behaviour works on the maintainer's hardware but regressions land all the time, the API and config surface shifts without notice, and a fresh deploy on someone else's gear has not been verified. Treat this as a hobby-grade work-in-progress, not a polished product. Bugs, PRs, and "this didn't work for me" issues all very welcome. 🍺☕ If you do try a fresh end-to-end deploy, please get in touch — I'll buy you a beer or a coffee. The best place to ask questions, get help, or show off a build is the [Dotty community Discord](https://discord.gg/7sKE5c6A).
 >
 > **Known rough edges:** face emoji rendering is missing visual differentiation for 4 of 9 emotions (sad / surprise / love / laughing); sound-direction localizer has a hardware-AEC-related left-bias on M5Stack CoreS3 (energy detection works, direction is unreliable); kid-voice ASR accuracy on SenseVoice has a kid-speech gap that whisper.cpp will close in a follow-up.
diff --git a/dotty-behaviour/README.md b/dotty-behaviour/README.md
index 399d735..8fc6931 100644
--- a/dotty-behaviour/README.md
+++ b/dotty-behaviour/README.md
@@ -35,6 +35,23 @@ ssh root@<UNRAID_HOST> '
 '
 ```
 
+### Vision-key env var (issue #15)
+
+Photo intents need an OpenAI-compatible API key for the VLM call. The
+compose file picks up any of these from the shell that runs
+`docker compose up`:
+
+- `VLM_API_KEY` (preferred)
+- `VISION_API_KEY` (fallback)
+- `OPENROUTER_API_KEY` (fallback of fallback)
+
+If none are set the container still starts, but `dispatch/vlm.py`
+returns the `VLM_OFFLINE_SENTINEL` string for every photo intent so
+the downstream LLM is told the camera is unavailable rather than
+confabulating a description. Set the key in the host shell before
+`docker compose up`, or pass `--env-file <path>` at a `.env` that
+contains it.
+
 ## Why a separate container
 
 The bridge was a separate process on the RPi for the whole life of
diff --git a/dotty-behaviour/docker-compose.yml b/dotty-behaviour/docker-compose.yml
index 7b0d419..38cf3af 100644
--- a/dotty-behaviour/docker-compose.yml
+++ b/dotty-behaviour/docker-compose.yml
@@ -39,6 +39,17 @@ services:
       - DOTTY_STATE_DIR=/var/lib/dotty-behaviour/state
       - HOUSEHOLD_YAML_PATH=/var/lib/dotty-behaviour/state/household.yaml
       - GREETER_STATE_PATH=/var/lib/dotty-behaviour/state/greeter_state.json
+      # Vision-language model credentials (issue #15). Required for
+      # photo intents — without these the container falls through to
+      # the VLM_OFFLINE_SENTINEL contract in dispatch/vlm.py and the
+      # downstream LLM is told the camera is unavailable rather than
+      # confabulating a description. Resolved in fallback order:
+      # VLM_API_KEY → VISION_API_KEY → OPENROUTER_API_KEY. Set any one
+      # in your shell or in .env next to this compose file.
+      - OPENROUTER_API_KEY=${OPENROUTER_API_KEY:-}
+      - VISION_API_KEY=${VISION_API_KEY:-}
+      - VLM_API_KEY=${VLM_API_KEY:-}
+      - AUDIO_CAPTION_API_KEY=${AUDIO_CAPTION_API_KEY:-}
     volumes:
       - /mnt/user/appdata/dotty-behaviour/state:/var/lib/dotty-behaviour/state
       - /mnt/user/appdata/dotty-behaviour/logs:/var/lib/dotty-behaviour/logs