Translation workbench

A two-pane translation workspace in the dashboard — source and target side by side, with on-device, local, or cloud runtimes and streaming, layout-preserving output.

The translation workbench is a dedicated, side-by-side workspace for translating longer text. Unlike the quick floating-menu translate action, the workbench gives you a full source/target layout, a choice of runtime, and streaming output that preserves the paragraph structure of your source.

Opening it

Open it from the dashboard (nav → Translate), or from the chat header's Open translation workbench action.

Layout

The workbench is a two-pane grid:

  • Source (left) — an editable text box where you paste or type what you want translated.
  • Target (right) — a read-only box where the translation streams in.

Above the output, Cordy shows the detected source language and a data-flow label telling you whether this translation is running On-device, on a Local model, or in the Cloud.

Choosing languages

Pick a Source and Target language from the selectors. Source includes Auto (detect automatically); the target must be a specific language. The available languages are English, 简体中文 (Simplified Chinese), 繁體中文 (Traditional Chinese), 日本語 (Japanese), 한국어 (Korean), Français, Deutsch, Español, Português, and Русский.

Use the swap button between the selectors to exchange source and target languages. When Source is Auto, the button becomes available after Cordy has detected a source language.

Runtimes

A three-way toggle chooses where the translation runs:

RuntimeWhat it usesData flow
DefaultChrome's built-in Translator API (on-device)On-device
LocalA browser-managed local model you've selected: Gemma (WebGPU) or Chrome's Gemini NanoLocal model
CloudA configured cloud provider and model (BYOK)Cloud

Default needs nothing beyond a supported Chrome. Local and Cloud require you to pick a model first — Cordy prompts you if you haven't. Local runtimes depend on local & on-device AI being set up; Cloud uses your provider keys. Gemini Nano translation in the workbench works since 2.5.7.

Translating

  1. Choose your runtime and languages.
  2. Paste or type into the Source pane.
  3. Click Translate.

The translation streams into the target pane and is revealed smoothly as it arrives; the pane follows the newest text until you scroll up. A status line shows Waiting for the model… Ns while the model prepares its answer, Thinking… while a reasoning model reasons, then Translating…, and finally Completed (or Stopped / Failed). The original paragraph and line-break structure is preserved.

How the source is sent depends on the runtime:

  • Cloud sends the whole text in one request whenever it fits the model's single-request budget. Cordy derives that budget from the model's context window and output limit — read at runtime from the provider's own model list, the Vercel AI Gateway catalog, or the public LiteLLM catalog — keeps about 10% headroom on both, and lets one request produce at most 65,536 output tokens. Under the source pane, Cordy shows the character count and an estimated token count.
  • If a cloud text is larger than the budget, Cordy asks before anything is sent: the dialog shows the estimated size, the model's limits and where they came from, and how many requests the text needs. Splits only fall between paragraphs, and the status line shows Part k of N. Cancel sends nothing.
  • Local translates paragraph by paragraph, and Default line by line, on your device.

Actions

ActionWhat it does
TranslateStart translating the source
StopHalt an in-progress translation (partial output is kept)
RetryRe-run the translation
ClearEmpty the panes to start over

Each pane has its own copy button: Copy source text on the source pane and Copy translation on the target pane (since 2.5.6).

Notes and limits

  • Layout preservation is plain-text. Cordy keeps paragraphs and line breaks, but the panes are plain text — Markdown, HTML, and rich formatting aren't reconstructed.
  • Local-native runtimes are excluded. The Local translation picker accepts browser-managed models (Gemma or Gemini Nano); Ollama, LM Studio, and other OpenAI-compatible local runtimes are not available for translation in this version.
  • Local and Cloud need a model. Select a local or cloud translation model before translating with those runtimes.
  • Leaving the page stops the translation. Closing or reloading the workbench stops a running translation, so nothing keeps running — or billing — in the background.
  • Output limit. If a model stops at its output limit, the text translated so far is kept and marked [Stopped at the model's output limit]. Shorten the text or pick a model with a larger output limit.
  • Unknown models. When no catalog lists your model, Cordy plans with a conservative assumed limit and says so in the split dialog.

Privacy

Where your text goes depends on the runtime you pick:

  • Default (on-device) and Local never leave your device.
  • Cloud sends your source text to the provider you configured — and the first cloud translation is covered by the one-time consent dialog.
  • To size requests, Cordy downloads two public model catalogs about once a day; those downloads carry none of your text or data (see the privacy policy).