misapad

A minimal frontend for pair-writing with LLMs, inspired by Mikupad, but updated for use with chat-completion models.

A single text buffer holds both your writing and the model's. Write a paragraph, hit Ctrl+Enter to have the model continue, edit the result, keep writing, continue again. Model-generated text is tinted, marking who's responsible for what. Bad gens are quickly removed with a Ctrl+Z.

Quickstart

npm install
npm run dev          # http://localhost:5173

Open settings, set a base URL, API key, and model. That's it!

Documents, settings, and API keys are stored in localStorage.

Keybinds

Key	Action
Ctrl+Enter	Continue from the cursor (ask or prefill mode, per settings)
Ctrl+Shift+Enter	Undo the last generation and generate its replacement
Ctrl+K	Inline instruction popup: insert at the cursor or rewrite a selection
Esc	Stop a running generation or close popup
Ctrl+Z / Ctrl+Shift+Z	Undo/redo generations

On touch devices, the same actions live in a floating bar above the keyboard (undo, redo, continue/stop, reroll, instruct). Desktop remains the primary/recommended platform.

Continue modes

While chat models can't do raw text completion natively, misapad offers two ways to approximate it:

• ask mode prompts the model to continue your text and output just the completion. Works with every chat model. Quality depends on instruction-following.
• prefill mode sends your document as a partial assistant message the model continues verbatim. Where supported, this is the closest thing to true text completion. Flavors:
  • prefix: true on the trailing assistant message: DeepSeek (base URL https://api.deepseek.com/beta) and Mistral.
  • continue_final_message: true: vLLM, TabbyAPI.
  • raw trailing assistant message: OpenRouter and other servers that continue it implicitly.

The Ctrl+K popup sends the whole document (budgeted) with an <INSERT_HERE/> marker at your cursor, or your selection wrapped in <REWRITE> tags, plus your instruction, and strips echoed context from the reply.

Providers (CORS)

Since calls come straight from the browser, the provider must allow cross-origin requests. Most endpoints do by default, but vLLM may need --allowed-origins '["*"]'.

If a provider does block browser origins, uncomment the proxy block in vite.config.ts and point the base URL at /api/v1.

Development

npm test             # unit + headless end-to-end tests (no inference)
npm run mock         # canned SSE server on http://localhost:11435/v1 — any model name works;
                     # ?delay=900 for slow tokens, API key "fail-auth" for the error path
npm run build        # static site in dist/

With the dev server and mock server both running, node tools/e2e.mjs drives a headless Chromium through the whole workflow (continue, undo/redo, cancel, popup, error toast, autosave, sessions) and prints PASS/FAIL per step. It needs a Chromium for Playwright (npx playwright install chromium, or point CHROME_BIN at one) and it clears the app's localStorage on :5173, so don't aim it at a profile with real writing.