Agent Panel

A VS Code extension that hosts a TUI agent CLI as a dedicated tab in the bottom panel, next to Terminal, Problems, and Output. One panel slot, any agent — switch between them on demand without opening more terminals.

Out of the box it knows about Claude Code and OpenCode. You can add or override agents via settings, so any CLI that runs as an interactive TUI works.

Features

• Bottom-panel tab that hosts an agent in a real PTY (xterm.js + node-pty).
• Built-in support for Claude Code and OpenCode; switch between them from a title-bar button.
• Custom agents via a single settings array — no rebuild required.
• Active agent persists per workspace and across workspaces (global fallback).
• Theme- and font-synced with VS Code's terminal.

Requirements

• VS Code 1.85 or later.
• The agent's CLI must be on PATH (or specified as an absolute path in settings). For the built-ins, that means claude and/or opencode.
• On Windows only: a system-installed Node.js. The PTY host runs out of process to dodge a ConPTY deadlock in Electron's bundled Node. macOS, Linux, and WSL use VS Code's bundled Node directly — no extra install needed.

WSL works the same as native Linux: the VS Code Server extension host runs inside the distribution, so process.platform is linux and the non-Windows code path applies.

Switching agents

Open the bottom panel and click the Agent Panel tab. Use the ↔ Switch Agent button in the view's title bar to pick from the available agents. The active agent's name appears next to the view title; your choice is remembered for the workspace.

The same picker is available from the command palette as Agent Panel: Switch Agent, alongside Restart Agent and Clear Agent Terminal.

Configuration

Adding a custom agent

// settings.json
"agentPanel.agents": [
  {
    "id": "aider",
    "label": "Aider",
    "command": "aider",
    "args": ["--no-auto-commits"],
    "env": { "AIDER_NO_TELEMETRY": "1" },
    "icon": "sparkle"
  }
]

command accepts either a bare name (resolved via where/which at spawn time) or an absolute path. args and env are optional. icon is an optional codicon id shown in the picker.

In env values, ~ (home directory) and ${env:VAR} are expanded at spawn time, so ~/.foo and ${env:HOME}/.foo both work cross-platform.

Multiple Claude Code profiles (work / personal)

Claude Code reads its config (credentials, memory, settings) from CLAUDE_CONFIG_DIR, falling back to ~/.claude. Define one custom agent per profile so each runs against an isolated config directory:

"agentPanel.agents": [
  {
    "id": "claude-work",
    "label": "Claude Code (Work)",
    "command": "claude",
    "icon": "sparkle",
    "env": { "CLAUDE_CONFIG_DIR": "~/.claude-work" }
  },
  {
    "id": "claude-personal",
    "label": "Claude Code (Personal)",
    "command": "claude",
    "icon": "sparkle",
    "env": { "CLAUDE_CONFIG_DIR": "~/.claude-personal" }
  }
]

The built-in claude-code entry stays available and uses Claude's default ~/.claude directory. Use the Switch Agent button to move between profiles; the active selection persists per workspace, so each workspace can default to a different profile if you like.

The first time you switch into a new profile, run /login inside Claude Code to authenticate against that config dir — credentials, memory, and settings are kept separate from then on.

Example: hide Claude Code, keep OpenCode

"agentPanel.hiddenBuiltins": ["claude-code"],
"agentPanel.defaultAgent": "opencode"

Installing in a remote (WSL, SSH, Dev Containers)

The extension declares extensionKind: ["workspace"], so it runs in the workspace extension host — i.e. inside the WSL distro / SSH remote / dev container, not on your local machine. That's where the agent CLI gets spawned and where the agent's working directory lives.

Three ways to install it remotely:

1. From the marketplace (once published): connect to the remote, open the Extensions view, search for Agent Panel, and click Install in WSL: <distro> (or the equivalent for your remote).
2. From a local install: install the extension locally first, then connect to your remote — VS Code prompts you to install it on the remote side too.
3. Sideloading a .vsix without publishing: build a package (npx @vscode/vsce package) and from a terminal inside the remote run code --install-extension /path/to/agent-panel-0.2.0.vsix. Or use the Extensions view's ⋯ → Install from VSIX... while focused on the remote.

For native Windows or native Linux/macOS (no remote), there's nothing special to do — the extension installs and runs in the local host.

Development

npm install
npm run build      # one-shot
npm run watch      # rebuild on save

Press F5 in VS Code with this folder open to launch the Extension Development Host. The included .vscode/launch.json is wired up for it. To run from the command line instead:

code --extensionDevelopmentPath=C:/src/Panel <some-test-workspace>

(Use forward slashes on Windows — Git Bash strips unquoted backslashes from arguments.)

Project layout

src/
  extension.ts              activation, command registration, QuickPick
  AgentRegistry.ts          built-ins + settings merge
  AgentTerminalProvider.ts  webview lifecycle, PTY host management
  ptyHost.js                child-process Node host that owns node-pty
  utils.ts                  locates VS Code's bundled node-pty
media/
  main.js                   xterm.js client running in the webview

The PTY host runs as a separate process. On Windows it is spawned with a system Node binary to dodge a ConPTY deadlock that hits Electron's bundled Node; on macOS, Linux, and WSL it uses Electron's Node directly, whose ABI matches VS Code's bundled node-pty. Either way, the host re-uses node-pty shipped inside VS Code's app.asar.unpacked, so the extension itself does not bundle native modules.

Caveats

• Renaming an agent in agentPanel.agents while a session is running does not auto-restart it; an info message asks you to restart manually. Use the Restart button.
• The extension reuses node-pty from VS Code's installation. A future VS Code refactor that moves or removes that module would break the agent spawn step.
• Versions before 0.2.0 used different command and view ids (claudeCodePanel.*). Any keybindings or pinned panel positions tied to those will need to be re-pointed at agentPanel.*.