Agent IDE
A Claude-Code-style AI agent embedded in VS Code / Cursor — it reads, edits, and runs commands against your real workspace. Runs on your Claude subscription (no API key needed). And you can drive it from your phone or Telegram.
Not a fork, not a chat box bolted on — a full agent in a side panel, with the editor-grade controls you actually need: live diffs, checkpoints, context management, parallel agents, and one-tap remote access.
✨ What it does
🤖 An agent that works in your repo
- Reads, edits, and runs commands in your local workspace via the Claude Agent SDK — exactly like Claude Code, driven from a panel.
- Terminal-style tool cards: every command shows its
$ cmd and live output; approvals are merged right into the card (press Enter to Allow).
- Modes per chat: Agent (full), Ask (read-only — answers without touching files), Plan (proposes a plan first), Auto-accept (runs edits and commands without prompting).
🪟 Parallel chats
- Run multiple agents at once, each in its own chat — start a refactor in one, research in another. One run per chat, switch freely; background chats keep working and catch you up when you return.
🔍 See and control every change
- Live diffs for everything the agent touches — review inline, Keep or Revert per file, or all at once.
- Checkpoints & Rewind: roll the conversation back to any earlier message, optionally reverting the files that turn changed.
🧠 Context management
- Per-chat context-window meter over the composer — always know how much you're sending.
- /compact to summarize the chat and reclaim context when it fills up.
- Auto-context: active file + selection (and optionally open tabs / problems) are attached automatically, so "fix this" just works.
- @-mention workspace files, paste or attach images, and drag-drop files into the chat.
📱 Remote control — from your phone & Telegram ⭐
- One tap to go remote: a secure public tunnel opens automatically (zero config —
cloudflared is fetched for you, no domain, no server to run).
- A Telegram bot lets you navigate your open folders → chats → and open a web Mini App with the exact same UI, 1:1.
- Live bidirectional sync: type from your phone and it appears on your desktop instantly — and vice-versa.
- Cross-device approvals: an Allow/Deny raised on the desktop can be answered from your phone (and the card closes everywhere).
- Upload photos from your phone, draw/annotate screenshots, and send them to the agent.
- Eyes — headless-browser screenshot + rendered DOM, so the agent can see your web app.
- Hands — a managed dev server it can start, tail logs from, and stop.
- Memory — a persistent project notebook the agent reads/writes.
- Scale — spawn read-only subagents for parallel research.
- Orchestrator — kick off a background multi-agent run from a goal.
⚙️ More
- Plan-usage gauges (5h / 7d limits) in the header.
- Commit with an AI-written message (
Agent IDE: Commit Changes).
- Hybrid engine — point
ANTHROPIC_BASE_URL at a local model (LM Studio / Ollama) and run offline.
- Status-bar HUD — live state, active model, and running session cost.
⚡ Install
Prereq: the Claude CLI logged in once — run claude login (uses your Claude subscription, no API key needed). An ANTHROPIC_API_KEY works too.
- Install the extension (Marketplace, or Install from VSIX… with a release build).
- Reload Window, then click the robot icon in the side bar.
- Type a request and Send. To go remote, press the 🛰 button in the panel header.
Remote (optional): to use the Telegram bot, set agentIde.botToken (from @BotFather) and agentIde.telegramUserId (from @userinfobot) — the extension will ask for them the first time you enable remote.
🔑 Key settings (agentIde.*)
| Setting |
What |
model |
Default model (the picker is populated live from your account). |
includeEditorContext |
Auto-attach the active file + selection to each message. |
autoFixDiagnostics |
After edits, feed the agent any new errors to fix (bounded). |
superTools |
Master switch for the experimental Eyes/Hands/Memory/Scale tools. |
botToken / telegramUserId |
Telegram remote access. |
baseUrl |
Point at a local-model proxy (LM Studio / Ollama) for offline use. |
Full list in Settings → Extensions → Agent IDE.
⌨️ Shortcuts
Ctrl/Cmd+Alt+A — open panel
Ctrl/Cmd+Alt+N — new chat
Ctrl/Cmd+Alt+K — add the current selection to the chat
- Enter on an approval card — Allow
Agent IDE is proprietary software. © 2026 — all rights reserved.
| |