Usage, images, and voice

Conversation usage

When providers return usage events, the composer can show cumulative input/output tokens, cache activity, estimated cost, and current context fraction. Unknown values remain unknown rather than being guessed.

These figures are operational estimates, not invoices. Provider accounting, cache rules, discounts, and subscription terms remain authoritative.

Subscription windows

Nerve currently retrieves subscription usage snapshots only for OAuth Anthropic and OpenAI Codex. Refresh attempts are debounced per provider. This is separate from cumulative conversation usage and does not track general API-key billing.

Image input

Custom model definitions can declare image input and the harness can pass image content. Pasted images become temporary local paths rather than durable attachments.

For text-only models, choose an image-capable model for explain_image under Settings → Tools, then enable the tool. The agent can explicitly request a detailed description of a JPEG, PNG, GIF, or WebP path. The tool is not disclosed to the model while disabled and is automatically absent when the primary model already supports images.

The configured vision provider receives the image bytes. A local OpenAI-compatible vision endpoint can keep processing local; cloud-provider usage and billing remain separate from the primary model. See Images and voice for clipboard behavior.

Image generation

Enable and configure generate_image under Settings → Tools → Image generation. The agent supplies a prompt; provider, model, and output defaults remain under user control. Generated raster files are retained as tool-call artifacts and linked from the transcript.

The initial OpenAI Codex integration uses ChatGPT subscription image access and limits, not OpenAI API-key billing. Because that endpoint is not a public OpenAI API contract, availability and accepted options can vary by account.

Voice input

Voice capture happens in the browser/Electron renderer, then uploads to ChatGPT’s transcription endpoint. It requires OpenAI Codex OAuth with an account ID and uses gpt-4o-transcribe.

Limits are 8 minutes and 25 MB, with three client retry attempts. The transcript returns as editable text; audio is not added to the conversation as an attachment.