Agent IDE
Claude-Code-style agent inside VS Code / Cursor, running on your Claude subscription.
⚡ Install
Prereqs: VS Code or Cursor installed, and the Claude CLI logged in — run claude login once (uses your Claude subscription, no API key needed).
Easiest — one command
With the repo cloned and GitHub CLI authenticated (gh auth login):
- Windows:
./install.ps1
- macOS / Linux:
./install.sh
The script grabs the latest packaged .vsix from Releases and installs it into VS Code/Cursor.
Manual
- Download the latest
.vsix from Releases.
- Install it either way:
- CLI:
code --install-extension agent-ide-x.y.z.vsix (or cursor --install-extension …)
- UI: Extensions panel → ⋯ menu → Install from VSIX… → pick the file.
- Reload Window (
Ctrl+R), then open the robot icon in the side bar.
A Claude-Code-style agent embedded in VS Code as an extension (not a fork).
The agent reads, edits, and runs commands against your local workspace via
the Claude Agent SDK — exactly like Claude Code, driven from a side panel.
Architecture (first vertical slice)
VS Code ──▶ extension host (CommonJS)
│ WebviewView panel (media/main.js, style.css)
│ ▲ events / approvals
▼ │
src/agent.ts ──▶ @anthropic-ai/claude-agent-sdk query()
│ └─ edits LOCAL files, runs shell
│ env: ANTHROPIC_BASE_URL ◀── transport seam
▼
Anthropic API (later: proxy → LM Studio)
- Read-only tools (Read/Glob/Grep/…) run automatically.
- Mutating tools (Edit/Write/Bash/…) go through an approval gate rendered
in the panel — Allow / Deny per call.
- The model, API key, and
ANTHROPIC_BASE_URL are extension settings
(agentIde.*). The base-URL setting is the seam for local models later.
Develop / run
npm install
npm run build # or: npm run watch
Then open this folder in VS Code and press F5 (Run Extension). In the new
window: open a project folder, click the robot icon in the activity bar, type a
request, Send.
Auth: set agentIde.apiKey, export ANTHROPIC_API_KEY, or rely on an existing
claude login.
Smoke-test the engine (no VS Code needed)
node smoke.mjs
Creates a temp dir and asks the agent to write hello.txt there (only the
Write tool is permitted). Prints whether the file was created — proves the
SDK edits local files in this environment.
Extras beyond the chat panel, all driven from the Command Palette (and a couple
of keybindings):
- Status-bar HUD — live agent state (idle / working), the active model, and
the running session cost. Click it to open the panel. Toggle with
agentIde.showStatusBar.
- Export Conversation to Markdown (
Agent IDE: Export Conversation to Markdown) — write the active chat to .agent-ide/exports/ as a clean
transcript for PRs, sharing, or archiving.
- Restore Checkpoint (
Agent IDE: Restore Checkpoint…) — pick any per-turn
snapshot and roll the workspace back to before that turn's edits.
- Session Usage (
Agent IDE: Show Session Usage) — cumulative cost + turn
count, with a reset.
- Keybindings —
Ctrl/Cmd+Alt+A open panel · Ctrl/Cmd+Alt+N new chat.
Next
- Translation proxy so
ANTHROPIC_BASE_URL can point at LM Studio (local models).
@-mention workspace files in the composer.
- Per-turn token/context-window meter in the panel.