Photos & voice

Some people mostly send photos, others mostly talk in voice messages. iRoof.top keeps both habits.

Understanding them

In the media library, Describe with AI describes stickers, GIFs and photos and transcribes voice messages with the media model, which must accept images and audio. Set it in Settings or with OPENAI_MEDIA_MODEL, for example google/gemini-3.5-flash-lite on OpenRouter. Descriptions are only made when you press the button, and with a local model and Offline mode the photos never leave your computer.

<voice message 0:05: «what was said»>

When one kind dominates, the style profile says so, e.g. "talks mostly in voice messages (10 of 10)".

Big libraries

A long chat can hold tens of thousands of stickers, GIFs and photos. The media library helps you spend the model's tokens where they matter:

Control	What it does
Kind tabs	All, Missing, Described, Stickers, GIFs, Photos, Videos, Voice, with counts
Sort	most used (default), least used, recently used, earliest used
Use count	hide items used fewer than 2, 3, 5, 10 or 25 times
Search	in descriptions and ids

Describe N with AI describes exactly what is shown and not described yet, top to bottom in the chosen order, four at a time. Each description is saved as soon as it arrives. The progress line shows how many are done, failures, the rate and the time left. The same button turns into Stop: stop at any time, and the next run continues with the rest. A good start for a big chat: Most used first and Used 3+ times.

Your voice messages

With an empty composer the send button becomes a microphone, as in Telegram. The recording is converted to WAV in the browser, uploaded and transcribed. Visitors of share links can record too.

INFO

Browsers allow recording only on https:// or localhost.

The persona's photos and voice messages

Connect an MCP server that makes images and one that speaks:

{
  "img": {"command": "npx", "args": ["-y", "@dtst/img"],
          "env": {"OPENAI_BASE_URL": "https://openrouter.ai/api/v1", "OPENAI_API_KEY": "${OPENAI_API_KEY}", "OPENAI_MODEL": "meta/muse-image"}},
  "tts": {"command": "npx", "args": ["-y", "@dtst/tts"],
          "env": {"OPENAI_BASE_URL": "https://openrouter.ai/api/v1", "OPENAI_API_KEY": "${OPENAI_API_KEY}",
                  "OPENAI_MODEL": "google/gemini-3.8-flash-lite-tts", "OPENAI_VOICE": "Laomedeia"}}
}

The persona then gets two more actions:

Action	What it does
send_voice {text}	calls the speech tool (WAV, retried as mp3) and sends a voice message
send_photo {prompt, caption?}	calls the image tool and sends a photo

• With a speech tool connected, the persona's own transcribed voice messages are replayed as send_voice, so the model sees the habit in the form it can continue.
• Voice in Chat details (for example Puck) gives the persona their own voice.
• Any other image or audio an MCP tool returns becomes chat media gen_… (checked by content, up to 25 MB, deleted with the chat).

Costs

Each photo or voice message is an extra paid call: about $0.01 per image with meta/muse-image, and a few seconds of TTS.