Add images and voice
Paste an image
Section titled “Paste an image”Copy an image and paste with Ctrl/Cmd+V inside the composer. Nerve intercepts clipboard files with an image/* MIME type, stores each image under the operating system temp directory in nerve/, and inserts newline-separated local paths at the cursor.
Accepted MIME families include PNG, JPEG, GIF, WebP, SVG, BMP, TIFF, and AVIF. The path lets the agent’s file-reading pipeline send image content to a capable model.
For a text-only model, configure and enable explain_image under Settings → Tools. The agent can then send the temporary path to the selected vision model and receive a detailed text explanation. The tool is completely hidden from vision-capable models and whenever it is disabled.
Dropping an image, another file, or a folder onto the desktop composer is different from clipboard image paste: it inserts the item’s existing filesystem path without copying, uploading, or creating a durable attachment. See Use the composer for the full workflow. In a browser or installed PWA, use @ completion to reference paths inside the project.
Generate an image
Section titled “Generate an image”Agents can use the provider-neutral generate_image tool to generate raster images. Each generated image is retained as a tool-call artifact, and its file path appears as a link in the conversation. The agent receives that path rather than inline image bytes.
Choose the provider, model, and provider-specific output defaults under Settings → Tools → Image generation. The initial OpenAI Codex provider defaults to gpt-image-2.5-flare, with gpt-image-2.5-sunburst available for higher-quality or precision-focused work. The agent supplies only the image prompt.
The OpenAI Codex provider uses ChatGPT’s subscription image endpoint and usage limits, not OpenAI API billing. Availability and accepted options can change with the connected account because this endpoint is not part of the public OpenAI API contract.
Record voice
Section titled “Record voice”Use the microphone control or its keyboard shortcut. Recording can target the main composer or a user-question reply. Nerve shares one recording session across those surfaces.
- Maximum duration: 8 minutes.
- Maximum upload: 25 MB.
- Client transcription retries: up to 3.
- Right-click an active recording to cancel it.
Recording and transcription appear as an inline chip at the editor’s insertion point, including in user-question replies. The transcribed text replaces that chip (or is appended if the marker is no longer available); it is not sent automatically. The microphone has a distinct active-recording state, and Escape or right-click cancels recording.
Voice requires OpenAI Codex/ChatGPT OAuth and uploads audio to ChatGPT’s audio transcription endpoint using gpt-4o-transcribe. An OpenAI API key alone is not sufficient. Common WebM, MP4/M4A, WAV, MPEG/MP3, OGG, and FLAC recordings are accepted.