AI Chat
Use AI Chat to ask about the documentation, compare layouts, and work with the playground through application tools. Answers depend on the selected model; check calculations and proposed edits before relying on them. Imported documents and corpus text are untrusted reference data, not instructions for the assistant.
What it can do
Section titled “What it can do”AI Chat can change the selected layouts and analysis corpus, open and arrange panes, adjust Browse and analysis controls, edit a Browse metric column, inspect keys and chords, and work with Try Out practice and calibration controls. It can also read native analysis results, documentation, saved versions, and working documents. It can edit and validate drafts for layouts, metrics, metric views, corpora, alphabets, and mixtures; publish a layout preview; and save or delete library documents after approval. Workspace tools include workspace and pane-layout selection, dock arrangement, version history, comments, and saved Trace Examples.
The assistant loads relevant tools as needed and reads document or analysis details in bounded pages. It uses the same application calculations and validation as the visible playground. It can download a local library backup after you enable private-data sharing; only the download status and filename are returned to the provider. You still choose files, import a backup, answer firmware setup questions, and press the physical keys needed to calibrate your keyboard. Synthetic Try Out key events operate only the on-screen emulator and do not represent your physical typing.
Corpus selection uses the same Apply action as the corpus chooser. Personal corpus, alphabet, and mixture choices require private-data sharing, including their names. If a choice becomes unavailable after changing libraries, ask the assistant to discover the choices again.
To import a corpus, open its editor, choose Source file, configure Format, Sampling seed, Target graphemes, and Initial alphabet, then use Preview sample. Choosing the file alone does not sample or save it. Review the preview, then use Save sample in this browser. Preview again after choosing a different file or changing the format, seed, or target. After closing or explicitly reopening the editor, choose the original file again if you need another preview.
With the editor open, the assistant can run Preview sample on your chosen file, Check sample alphabet, Rebuild mixture, and the available export or cancel actions. It can export documents or unsaved recovery results, and request cancellation even when you started the preview or rebuild yourself. These actions require private-data sharing; previewing and rebuilding also require Accept edits. Coverage counts appear in the editor, and exports download locally. The actions return completion status without file contents or coverage counts. Saving remains separate and requires approval in built-in chat.
Canceling a preview or rebuild keeps the selected file and previous sample when interrupted before the new sample is attached. It cannot undo an attachment already completed. Closing the editor releases the selected file.
Corpus, alphabet, and mixture documents save in this browser, including while you are signed in. Keep a local library backup before clearing browser data. A saved corpus or mixture keeps its sampled text and the recipe used to produce it. Editing source settings alone does not resample that text or rebuild a mixture; Save requires a sample that matches the draft. Later edits to a mixture’s sources do not change its saved sample.
Provider setup
Section titled “Provider setup”Open Setup in AI Chat. Choose OpenRouter to connect your own OpenRouter account through its authorization page, or choose a local OpenAI-compatible endpoint. No application API key is supplied. The assistant never switches to a remote provider when a local connection fails.
For OpenRouter, finish authorization in the same browser tab and select a model that supports tool calls. Canceling authorization leaves the provider disconnected; retry Setup when ready. OpenRouter usage is charged to your account.
For a local provider, start a server such as LM Studio or Ollama on the same machine as your browser. Enter its OpenAI-compatible base URL, including /v1, for example http://localhost:1234/v1. Enter an API key only if that server requires one, and choose a model with tool-call support. A model without tool support may answer text questions but cannot reliably control the application.
Credentials remain in memory for the current page session. Disconnect cancels work and clears the active credentials and conversation. Disconnecting does not revoke an OpenRouter key at OpenRouter; revoke it separately in your account when needed. Do not paste keys into the conversation.
Permissions and sharing
Section titled “Permissions and sharing”The default mode reads documentation and application state. Navigation, opening panes, and changing layout, corpus, analysis, or filter selections are allowed within your sharing permissions. Accept edits additionally permits changes to unsaved drafts. These drafts can be autosaved locally for recovery; editing them is distinct from updating a saved library item. Saving and deleting always require human approval in built-in chat, including with Accept edits enabled. Replacing an existing Browse metric column also requires approval because its previous definition has no version-history restore. Review the requested action before approving it.
Private document and corpus names, corpus contents, words and n-grams, private drafts, and results derived from private data can reveal personal information. Even an aggregate can disclose a small or distinctive private dataset. Enable private-data sharing only when you intend to send that information to the selected provider. Saving a document, signing in, or accepting a storage notice does not enable AI sharing. Local inference sends it to the endpoint you entered; a remote endpoint sends it off your machine. Your own chat messages are sent to the selected provider, so do not type private information unless you intend to share it. Public documentation is available without private-data sharing.
The assistant reads details only when needed instead of receiving every document or corpus in its initial context. Each message can use up to four model requests, eight application actions, and 1,024 output tokens per model request. Use Cancel to stop a request. A canceled request may already have completed an application action; inspect the current state before retrying.
Local connection troubleshooting
Section titled “Local connection troubleshooting”Allow a browser local-network permission request only for the server you intend to use. Browser and operating-system support varies. If access is blocked or unsupported, check the browser’s site permissions and the server’s configuration. The endpoint is on the browser’s machine, which may differ from the machine hosting the website.
The local server must allow the exact website origin through CORS, including preflight OPTIONS, POST, and the Content-Type header. If a key is used, it must also allow the Authorization header. A failed preflight or missing permission cannot be repaired by changing the model. Keep the application’s cross-origin isolation protections enabled; local inference does not require weakening COOP or COEP. The application does not scan ports or create a proxy to expose your local server.
External agents
Section titled “External agents”Supported browsers can expose application tools to an external agent through WebMCP. Availability depends on the browser and agent integration; ordinary browser support does not guarantee tool discovery. External agents use their own permission controls. A normal document save adds a recoverable version to history; deletion can remove a saved item and its history. Actions that overwrite data without version history, such as replacing a saved calibration or pane preset, have a separate confirmation boundary. Review the exact action and any page confirmation carefully.