ASKII ( ••)>⌐■-■ (⌐■■)

An AI code assistant for VS Code with style — ask about your code, have it edited for you, run agentic tasks, drive your screen or a browser, and keep AI-classified notes, tasks and reminders. Bring your own LLM: ASKII Cloud, Ollama, LM Studio, OpenAI, Anthropic, or opencode Go.

Install: VS Code Marketplace · Open VSX · CLI on npm

Quick start

1. Install ASKII from the VS Code Marketplace — or build it from source (see Development)
2. Run ASKII: Open Setup (Ctrl+Alt+K S / Cmd+Alt+K S) — a short wizard that walks you through provider, models, options and wiki setup
3. Select some code and run Ask ASKII (Ctrl+Alt+K A)

Commands

Command	Keybinding	What it does
Ask ASKII	Ctrl+Alt+K A	Ask about the selected code (or anything) — answers stream into a markdown panel with follow-ups and a history view
ASKII Edit	Ctrl+Alt+K E	Replace the selection with an AI-edited version; review the diff and undo if needed
ASKII Do	Ctrl+Alt+K D	Agentic workspace agent — explores, creates, modifies, renames and deletes files over multiple rounds
ASKII Control	Ctrl+Alt+K C	Screen or browser agent — screenshot-driven mouse/keyboard control, or Puppeteer web automation
ASKII Note	Ctrl+Alt+K N	Notes / tasks / reminders panel with search and screenshots
ASKII: Reload Wiki	Ctrl+Alt+K W	Rebuild the wiki RAG index
ASKII: Clear Cache	Ctrl+Alt+K X	Clear cached inline explanations and completions
ASKII: Open Setup	Ctrl+Alt+K S	Re-run the setup wizard

On macOS swap Ctrl for Cmd. The (⌐■_■) status-bar button opens a quick-access menu for all of these.

Ask ASKII

Select some code and ask about it, or ask a general question with nothing selected — the file name, language and selection are included as context automatically (plus matching wiki chunks when wiki RAG is on). Answers stream into the markdown panel with syntax-highlighted code blocks, and the panel header gives you copy, follow-up (previous questions and answers are kept as conversation context) and a history view to re-read the whole session.

File citations are clickable: when the answer references a file or code location, references like src/app.ts:42 render as links — click one to open the file in the editor at exactly that line (ranges like src/app.ts:10-20 select the whole span). Works for workspace code and docs-wiki files alike.

ASKII Edit

Describe the change you want and ASKII rewrites your selection — or the whole file if nothing is selected, using the full file as context either way. The edit is applied immediately, then a side-by-side diff of original vs. proposed code opens beside it so you can review it, with a one-click Undo offer to revert. Optional askii.formatAfterEdit formats the document after the edit.

ASKII Do

Describe a task — ASKII prints the top-level workspace structure, then loops until the task is done or askii.doMaxRounds (default 20) is reached. Read-only actions (list, view, search, wiki_search) run freely; write actions (create, write, modify, rename, copy, mkdir, delete) and shell run ask for confirmation (skip with askii.doAutoConfirm). Files are backed up before destructive changes, and the whole run can be undone with one click when it finishes.

ASKII Control

Pick Screen or Browser, then describe the task. ASKII screenshots the current state each round and proposes the next action with its reasoning:

• Screen — mouse_move, mouse_left_click, mouse_right_click, keyboard_input, DONE
• Browser — goto, click, type, wait_for, back, forward, DONE (Puppeteer)

Requires a vision-capable model (e.g. llava, moondream2). Browser mode auto-detects Chrome / Edge / Chromium / Brave (or set askii.chromePath) and is visible by default (askii.browserHeadless hides it); the browser closes when the loop ends. Screen control uses shell commands — Linux needs xdotool.

ASKII Note

Type free text — the AI classifies it as a note, a task (low / medium / high priority) or a reminder (with a due time), and asks a clarifying question when the intent or time is ambiguous. Entries are stored globally, tagged by workspace, full-text searchable, and grouped by age (Pinned, Today, Yesterday, This week, Earlier). Hover an entry to pin, edit (with AI re-classify), mark done, snooze / dismiss, or delete it.

• Time triggers — a wide range of phrases resolve locally, right in the composer: in 30 minutes, in 1h30m, in half an hour, in two hours, in a couple of hours, in a few minutes, 20 minutes from now, an hour from now, at 5pm, at noon, tomorrow at noon, at 9 in the morning, half past 5, quarter to 6, later today, tonight at 8, tomorrow morning, day after tomorrow, this time tomorrow, first thing in the morning, on friday, friday evening, by tuesday at 5pm, next monday, friday next week, next week on wednesday, this weekend, end of day, end of month, end of year, on the 15th, by the 1st of next month, july 8th, 2026-07-08 14:30, …
• Recurring triggers — every day at 9am, daily, every morning, every weekday at 9, every monday at 10, every monday and wednesday at 6pm, every mon, wed, fri, weekly on friday at 5pm, every weekend, every 2 hours, every hour, every half hour, every other day, biweekly, quarterly, every month, every month on the 15th, every year on july 8 — the entry re-arms itself after each firing and shows a repeat badge
• Event triggers — fire on editor events instead of a clock time: when I open VS Code, on startup, when I open this workspace, when I open package.json, every time I save .env, when I open a ts file, when I save this file, or bare when I save. One-shot triggers fire once (shown with a ⚡ badge), every time/whenever triggers re-arm on each matching event
• Task deadlines & recurring tasks — tasks accept due times too (fix the login bug by friday) and can repeat (water the plants every 3 days); when a deadline passes you get a task-due notification with Open / Done / Snooze, and un-completing a done task re-arms it
• Reminders fire as VS Code notifications with a sound while VS Code is running; ones missed while it was closed re-fire on the next startup
• Screenshots — toggle the camera icon to attach a full-screen capture to the entry

Inline completion & decorations

• askii.inlineCompletionEnabled — Copilot-style ghost text; Tab accepts, Esc dismisses. Tune the trigger with askii.inlineCompletionEagerness (low 1200 ms / medium 500 ms / high 200 ms). Suggestions are cached per context, and matching context from your enabled wikis is injected into the prompt
• askii.inlineHelperMode — adds a kaomoji decoration after the current line: helpful, funny, or wiki (explanations enriched by matching context from your enabled wikis — docs and code). Example:

const sum = a + b; (◕‿◕) Adds two variables; prefer const for variables that won't be reassigned.

Wiki RAG

1. Point askii.wikiPath at a folder of .md docs
2. Run ASKII: Reload Wikis — chunks the docs by heading and builds a vector index (embeddings run locally via embeddinggemma2)
3. Enable askii.wikiEnabled — matching chunks are injected into Ask / Edit / Do as context
4. Optional: askii.wikiAutoReload rebuilds the index on startup

Code wiki

A vector index over your workspace code (code files only — no .md, .txt, .json, markup, or images) used as retrieval context for Ask / Edit / Do:

1. Run ASKII: Reload Wikis (or askii wiki-reload --use-code-wiki) — walks the workspace (respecting root .gitignore and skipping node_modules/build dirs), chunks code into overlapping 60-line windows, and embeds them
2. Enable askii.codeWikiEnabled — relevant code chunks are injected into Ask / Edit as context, and the Do agent gets a code_search action for semantic lookup
3. Optional: askii.codeWikiAutoReload rebuilds the index on startup

Both wikis run embeddinggemma2 fully locally in JavaScript (Transformers.js + ONNX WASM — no API keys, no external services). The model downloads once (~314 MB) to ~/.askii/models on first use and is shared by the CLI and the extension. Rebuilds are incremental (unchanged files keep their vectors) and the index is memory-optimized: Int8-quantized vectors (~¼ the RAM of Float32), metadata packed as typed arrays, and chunk text stays on disk — only the top hits are read back. Indexes live in .askii/code-wiki/ (workspace) and .askii-wiki-index/ (wiki folder).

Commit messages

A ✦ button in the Source Control view toolbar generates a commit message from your staged (or working-tree) diff straight into the input box. It styles the message after your repo's last 10 commits, honors askii.commitMessageStyle (oneliner / brief / descriptive), and appends your own rules from the .md file in askii.commitMessageInstructions. While it runs, the button turns into a ■ stop button that cancels it.

Code review

The Code Review Uncommitted Changes button (checklist icon) in the Source Control toolbar sends all uncommitted changes (staged, unstaged and untracked files) to the Ask / Edit / Do model. It opens a review panel that lists each problem with its severity, a file:line link and a recommendation. While the review runs, the button turns into a ■ stop button that cancels it.

• Auto Fix asks the model to fix that one problem and applies the fix to the file as an unsaved edit (Ctrl+Z undoes it)
• Fixed It marks a problem you fixed yourself; Reopen undoes the mark
• Show fixed hides or shows the problems that are already fixed

LLM platforms

Platform	Type	Setup
ASKII Cloud (default)	Cloud	API key in askii.askiicloudApiKey — base URL is fixed to https://api.askii.dev/v1
Ollama	Local	Install + pull a model (e.g. gemma4:e4b), runs at http://localhost:11434
LM Studio	Local	Enable the local server (ws://localhost:1234) and load a model
OpenAI or compatible	Cloud	askii.openaiApiKey; set askii.openaiUrl for Azure etc. (empty = api.openai.com)
Anthropic	Cloud	askii.anthropicApiKey
opencode Go	Cloud	askii.opencodegoApiKey

Each feature group picks its own platform and model, so you can mix providers:

Feature group	Platform	Model	Default model
Ask / Edit / Do, code review	askii.llmPlatform	askii.llmModel	askii-smart
Inline suggestions, completion, commit messages	askii.llmInlinePlatform	askii.llmInlineModel	askii-fast
Control / Note (vision)	askii.llmVisionPlatform	askii.llmVisionModel	askii-smart

API keys and server URLs are shared per provider across all groups.

Configuration

All settings live in VS Code Settings under the askii.* namespace.

Notes & reminders

Setting	Default	Description
askii.noteReminderSound	true	Play a sound when a reminder fires
askii.noteSnoozeMinutes	10	Minutes added by Snooze

Agents & browser

Setting	Default	Description
askii.doMaxRounds	20	Max rounds for Do / Control
askii.doAutoConfirm	false	Skip write confirmations
askii.formatAfterEdit	false	Format files after Edit / Do
askii.browserHeadless	false	Hide the Puppeteer browser window
askii.chromePath	(empty)	Browser executable (empty = auto-detect)

Git

Setting	Default	Description
askii.commitMessageStyle	oneliner	Generated message style
askii.commitMessageInstructions	(empty)	.md file with extra commit rules

Wiki RAG

Setting	Default	Description
askii.wikiEnabled	false	Inject docs wiki context into Ask / Edit / Do
askii.wikiPath	(empty)	Folder of .md docs
askii.wikiAutoReload	false	Rebuild docs wiki on startup
askii.codeWikiEnabled	false	Inject code wiki context into Ask / Edit / Do
askii.codeWikiAutoReload	false	Rebuild code wiki on startup

Embeddings for both wikis run locally (embeddinggemma2) — no embedding settings required.

CLI

ASKII also ships as a terminal CLI with an interactive REPL and the same ask / edit / do / control / browse / note commands:

npm install -g askii-cli

See cli/README.md.

Requirements

• VS Code 1.108+
• Control requires a vision-capable model; screen control needs xdotool on Linux; browser mode needs a Chromium-based browser

Development

git clone https://github.com/danisss9/ASKII.git
cd ASKII && npm install

npm run watch     # tsc + esbuild watchers — then F5 to launch the extension
npm run compile   # type-check + lint + bundle
npm run package   # production bundle (for vsce)
npm test          # integration tests (@vscode/test-electron)

• src/ — extension host code
• common/ — pure Node code shared with the CLI (never imports vscode)
• cli/ — the CLI package (npm run build / npm run dev)

License

MIT