PiAgent - AI Coding Agent for VSCode

A full-featured AI coding agent that lives inside VSCode's Chat panel. Powered by pi-coding-agent — the same engine behind the pi CLI. Mention @piagent and let it read, write, edit files and execute bash commands to complete coding tasks autonomously.

Note: All features below — multi-provider support, tool execution, session management, OAuth login, auto-compaction, and more — are provided by the pi-coding-agent library. This extension is a thin VSCode integration layer on top of that powerful foundation.

Why PiAgent?

• Not locked to one model. Use Claude, GPT-4, Gemini, DeepSeek, Mistral, or any of 15+ supported LLM providers. Switch models mid-conversation.
• Real tools, not autocomplete. The agent reads files, writes files, makes surgical edits, and runs shell commands — the same four tools the pi CLI uses.
• No permission prompts. There is no permission system. The agent executes tools directly without asking for approval on every file read or shell command. This is a deliberate design choice — it keeps the workflow fast and uninterrupted. You are in control of what you ask it to do.
• Shared configuration with the pi CLI. API keys, settings, models, and auth are stored in ~/.pi/agent/, not in VSCode settings. If you already use the pi CLI, PiAgent works out of the box with the same config. Nothing is VSCode-specific.
• Session persistence. Start a session, close VSCode, resume it later. Sessions are stored on disk and shared with the pi CLI.
• Automatic context compaction. When the conversation gets too long, PiAgent compacts the context automatically so you can keep working without hitting token limits.

Quick Start

1. Install the extension from the VSCode Marketplace

2. Authenticate with a provider — pick one:

   Option A: OAuth login (recommended for subscriptions)

   Open the Chat panel, type @piagent /login, and select your provider. PiAgent opens your browser to complete authentication. Supports Anthropic (Claude Pro/Max), OpenAI (ChatGPT Plus/Pro), GitHub Copilot, and Google (Gemini CLI / Antigravity).

   Option B: API key via environment variable

   export ANTHROPIC_API_KEY=sk-ant-...
   # or
   export OPENAI_API_KEY=sk-...
   # or
   export GEMINI_API_KEY=...
   

   Option C: API key in auth file

   Store keys in ~/.pi/agent/auth.json (same file the pi CLI uses).

3. Open VSCode's Chat panel (Ctrl+Shift+I / Cmd+Shift+I)

4. Type @piagent followed by your prompt

That's it. PiAgent initializes a session, picks the best available model, and starts working.

Features

Chat Participant

Mention @piagent in VSCode's built-in Chat panel. PiAgent registers as a native chat participant — no separate webview, no sidecar process.

Tool Execution

PiAgent has four tools:

All tool output streams to the Chat panel in real time. Full output is available in the Output channel (View → Output → PiAgent).

Multi-Provider LLM Support

PiAgent supports every provider that pi-coding-agent supports:

Switch models at any time with /model or Cmd+Shift+M.

OAuth Login

Use /login to authenticate with subscription-based providers directly from VSCode — no need to manually edit JSON files or set environment variables. PiAgent opens your browser, completes the OAuth flow, and saves credentials to ~/.pi/agent/auth.json.

Supported OAuth providers:

• Anthropic — Claude Pro / Max subscription
• OpenAI — ChatGPT Plus / Pro (Codex)
• GitHub Copilot — existing Copilot subscription
• Google Gemini CLI — Google Cloud authentication
• Google Antigravity — Gemini 3, Claude, GPT via Google Cloud

Use /logout to remove stored credentials for any provider.

Session Management

• /new — Start a fresh session
• /resume — Pick from previous sessions (with message count and timestamp)
• /session — View current session ID, model, message count, and working directory
• /compact — Manually compact context to free up token space

Sessions are stored on disk at ~/.pi/agent/sessions/ and are fully compatible with the pi CLI. You can start a session in the terminal and resume it in VSCode, or vice versa.

Automatic Context Compaction

When the conversation approaches the model's context window limit, PiAgent automatically compacts the history — summarizing earlier messages while preserving recent context. This happens transparently; you see a brief progress indicator and can keep working.

Slash Commands

Type @piagent / to see all available commands:

Command Palette

Open the command palette (Ctrl+Shift+P / Cmd+Shift+P) and search for:

Keyboard Shortcuts

Configuration

PiAgent shares most configuration with the pi CLI — API keys, models, and sessions are stored in ~/.pi/agent/, not in VSCode settings. This means your setup works identically whether you use PiAgent in VSCode, the pi CLI in a terminal, or any other tool built on pi-coding-agent.

VSCode Settings

PiAgent has VSCode-specific settings for status bar display and agent behavior. Use /settings in chat or open VSCode settings and search for "piagent".

piagent.statusBar.show

Controls which token usage stats are shown in the status bar next to the model name. By default, everything is shown:

claude-opus-4-6 ↑141 / ↓26k / R7.8M / W99k / $5.159 (sub) / 49.7%/200k (auto)

Each item can be toggled independently:

Default — show everything:

"piagent.statusBar.show": [
  "inputTokens",
  "outputTokens",
  "cacheRead",
  "cacheWrite",
  "cost",
  "contextUsage"
]

Show only cost and context usage:

"piagent.statusBar.show": ["cost", "contextUsage"]

Show only the model name (hide all stats):

"piagent.statusBar.show": []

The tooltip (hover over the status bar) always shows the full breakdown regardless of this setting.

piagent.autoCompaction

Automatically compact the conversation context when approaching the model's token limit. When enabled (default: true), older messages are summarized to free up space while preserving recent context.

piagent.autoRetry

Automatically retry failed API requests with exponential backoff (default: true). Helps handle transient network errors and rate limits.

piagent.blockImages

Block image attachments from being sent to the model (default: false). Useful for reducing token usage or when using models that don't support vision.

piagent.thinkingLevel

Default thinking level for reasoning models like Claude with extended thinking or o1 (default: "medium"). Options: "off", "minimal", "low", "medium", "high". Higher levels allow more thorough reasoning but use more tokens.

API Keys

Set environment variables before launching VSCode, or store them in ~/.pi/agent/auth.json:

# Environment variables
export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...
export GEMINI_API_KEY=...
export DEEPSEEK_API_KEY=...
export MISTRAL_API_KEY=...
export GROQ_API_KEY=...
export XAI_API_KEY=...

Settings

Global settings live at ~/.pi/agent/settings.json. Project-local overrides go in .pi/settings.json at the project root.

Custom Models

Add or override models via ~/.pi/agent/models.json. Any provider that speaks the OpenAI, Anthropic, or Google API can be added as a custom provider. See the pi documentation for details.

Output

• Chat Panel — Streaming markdown responses and inline tool call summaries
• Output Channel — Full untruncated tool output, useful for long bash output or large file reads. Access via View → Output → PiAgent
• Status Bar — Shows the active model, token usage, cost, and context window usage. Click to switch models. Hover for a detailed breakdown. Customize visible items via piagent.statusBar.show.

Requirements

• VSCode 1.100.0 or later
• Node.js 20.0.0 or later
• An API key for at least one supported provider

FAQ

How is this different from GitHub Copilot?

PiAgent is an autonomous coding agent with file system access and shell execution. It can read your codebase, make multi-file edits, run tests, and iterate on errors. Copilot is primarily an autocomplete and chat tool. PiAgent also lets you bring any model from any provider.

Do I need the pi CLI installed?

No. PiAgent bundles pi-coding-agent as a dependency. However, if you already use the pi CLI, they share the same configuration directory (~/.pi/agent/), so API keys, settings, sessions, and custom models work in both places.

Can I use my Anthropic/OpenAI subscription instead of an API key?

Yes. Use /login directly in VSCode to authenticate with your existing subscription (Anthropic Claude Pro/Max, ChatGPT Plus/Pro, GitHub Copilot, or Google Gemini). PiAgent opens your browser, completes the OAuth flow, and saves credentials to ~/.pi/agent/auth.json. If you've already authenticated via pi /login in the CLI, PiAgent picks up the stored credentials automatically.

Does PiAgent ask for permission before running tools?

No. There is no permission system — no "allow/deny" dialogs for file edits or shell commands. The agent executes tools directly as requested. This matches the behavior of the pi CLI. If you want to review changes before they happen, ask the agent to show you a plan first, or use git to review and revert.

Where are sessions stored?

Sessions are stored at ~/.pi/agent/sessions/ and <project>/.pi/sessions/. They are plain JSON files and are fully portable between PiAgent and the pi CLI.

License

MIT

Credits

Built on pi-coding-agent and pi-ai by Mario Zechner. These libraries provide all the core functionality — multi-provider LLM support, tool execution, session management, OAuth authentication, and more. This extension is a VSCode integration layer that brings that power to the Chat panel.