co-mono/packages/coding-agent/README.md

pi

A terminal-based coding agent with multi-model support, mid-session model switching, and a simple CLI for headless coding tasks.

Works on Linux, macOS, and Windows (requires bash; see Windows Setup).

Table of Contents

• Getting Started
  • Installation
  • Windows Setup
  • API Keys
  • Quick Start
• Usage
  • Slash Commands
  • Editor Features
  • Keyboard Shortcuts
  • Bash Mode
  • Image Support
• Sessions
  • Session Management
  • Context Compaction
  • Branching
• Configuration
  • Project Context Files
  • Custom Models and Providers
  • Themes
  • Custom Slash Commands
  • Skills
  • Hooks(#hooks)
  • Settings File
• CLI Reference
• Tools
• Programmatic Usage
• Philosophy
• Development
• License

---

Getting Started

Installation

npm (recommended):

npm install -g @mariozechner/pi-coding-agent

Standalone binary:

Download from GitHub Releases:

Platform	Archive
macOS Apple Silicon	pi-darwin-arm64.tar.gz
macOS Intel	pi-darwin-x64.tar.gz
Linux x64	pi-linux-x64.tar.gz
Linux ARM64	pi-linux-arm64.tar.gz
Windows x64	pi-windows-x64.zip

# macOS/Linux
tar -xzf pi-darwin-arm64.tar.gz
./pi

# Windows
unzip pi-windows-x64.zip
pi.exe

macOS note: The binary is unsigned. If blocked, run: xattr -c ./pi

Build from source (requires Bun 1.0+):

git clone https://github.com/badlogic/pi-mono.git
cd pi-mono && npm install
cd packages/coding-agent && npm run build:binary
./dist/pi

Windows Setup

Pi requires a bash shell on Windows. Checked locations (in order):

1. Custom path from ~/.pi/agent/settings.json
2. Git Bash (C:\Program Files\Git\bin\bash.exe)
3. bash.exe on PATH (Cygwin, MSYS2, WSL)

For most users, Git for Windows is sufficient.

Custom shell path:

// ~/.pi/agent/settings.json
{
  "shellPath": "C:\\cygwin64\\bin\\bash.exe"
}

API Keys

Set the environment variable for your provider:

Provider	Environment Variable
Anthropic	ANTHROPIC_API_KEY or ANTHROPIC_OAUTH_TOKEN
OpenAI	OPENAI_API_KEY
Google	GEMINI_API_KEY
Mistral	MISTRAL_API_KEY
Groq	GROQ_API_KEY
Cerebras	CEREBRAS_API_KEY
xAI	XAI_API_KEY
OpenRouter	OPENROUTER_API_KEY
ZAI	ZAI_API_KEY

The /model command only shows models for providers with configured API keys.

OAuth (Claude Pro/Max subscribers):

pi
/login  # Select "Anthropic (Claude Pro/Max)", authorize in browser

Tokens stored in ~/.pi/agent/oauth.json (mode 0600). Use /logout to clear.

GitHub Copilot:

pi
/login  # Select "GitHub Copilot", authorize in browser

During login, you'll be prompted for an enterprise domain. Press Enter to use github.com, or enter your GitHub Enterprise Server domain (e.g., github.mycompany.com). All models are automatically enabled after login.

If you get "The requested model is not supported" error, enable the model manually in VS Code: open Copilot Chat, click the model selector, select the model, and click "Enable".

For enterprise users, check with your organization's Copilot administrator for model availability.

Tokens stored in ~/.pi/agent/oauth.json (mode 0600). Use /logout to clear.

Quick Start

export ANTHROPIC_API_KEY=sk-ant-...
pi

Then chat:

You: Create a simple Express server in src/server.ts

The agent reads, writes, and edits files, and executes commands via bash.

---

Usage

Slash Commands

Command	Description
/model	Switch models mid-session (fuzzy search, arrow keys, Enter to select)
/thinking	Adjust thinking level for reasoning models (off/minimal/low/medium/high)
/queue	Set message queue mode: one-at-a-time (default) or all-at-once
/export [file]	Export session to self-contained HTML
/session	Show session info: path, message counts, token usage, cost
/changelog	Display full version history
/branch	Create new conversation branch from a previous message
/resume	Switch to a different session (interactive selector)
/login	OAuth login for subscription-based models
/logout	Clear OAuth tokens
/clear	Clear context and start fresh session
/copy	Copy last agent message to clipboard
/compact [instructions]	Manually compact conversation context
/autocompact	Toggle automatic context compaction
/theme	Select color theme
/show-images	Toggle inline image display (supported terminals only)

Editor Features

File reference (@): Type @ to fuzzy-search project files. Respects .gitignore.

Path completion (Tab): Complete relative paths, ../, ~/, etc.

Drag & drop: Drag files from your file manager into the terminal.

Multi-line paste: Pasted content is collapsed to [paste #N <lines> lines] but sent in full.

Message queuing: Submit messages while the agent is working. They queue and process based on /queue mode. Press Escape to abort and restore queued messages to editor.

Keyboard Shortcuts

Navigation:

Key	Action
Arrow keys	Move cursor / browse history (Up when empty)
Option+Left/Right	Move by word
Ctrl+A / Home	Start of line
Ctrl+E / End	End of line

Editing:

Key	Action
Enter	Send message
Shift+Enter / Alt+Enter	New line (Ctrl+Enter on WSL)
Ctrl+W / Option+Backspace	Delete word backwards
Ctrl+U	Delete to start of line
Ctrl+K	Delete to end of line

Other:

Key	Action
Tab	Path completion / accept autocomplete
Escape	Cancel autocomplete / abort streaming
Ctrl+C	Clear editor (first) / exit (second)
Shift+Tab	Cycle thinking level
Ctrl+P	Cycle models (scoped by --models)
Ctrl+O	Toggle tool output expansion
Ctrl+T	Toggle thinking block visibility

Bash Mode

Prefix commands with ! to execute them and add output to context:

!ls -la
!git status
!cat package.json | jq '.dependencies'

Output streams in real-time. Press Escape to cancel. Large outputs truncate at 2000 lines / 50KB.

The output becomes part of your next prompt, formatted as:

Ran `ls -la`

``` ```

Run multiple commands before prompting; all outputs are included together.

Image Support

Attaching images: Include image paths in your message:

You: What's in this screenshot? /path/to/image.png

Supported formats: .jpg, .jpeg, .png, .gif, .webp

Inline rendering: On terminals that support the Kitty graphics protocol (Kitty, Ghostty, WezTerm) or iTerm2 inline images, images in tool output are rendered inline. On unsupported terminals, a text placeholder is shown instead.

Toggle inline images with /show-images or set terminal.showImages: false in settings.

---

Sessions

Session Management

Sessions auto-save to ~/.pi/agent/sessions/ organized by working directory.

pi --continue      # Continue most recent session
pi -c              # Short form

pi --resume        # Browse and select from past sessions
pi -r              # Short form

pi --no-session    # Ephemeral mode (don't save)

pi --session /path/to/file.jsonl  # Use specific session file

Context Compaction

Long sessions can exhaust context windows. Compaction summarizes older messages while keeping recent ones.

Manual: /compact or /compact Focus on the API changes

Automatic: Enable with /autocompact. When enabled, triggers in two cases:

• Overflow recovery: LLM returns context overflow error. Compacts and auto-retries.
• Threshold maintenance: Context exceeds contextWindow - reserveTokens after a successful turn. Compacts without retry.

When disabled, neither case triggers automatic compaction (use /compact manually if needed).

How it works:

1. Cut point calculated to keep ~20k tokens of recent messages
2. Messages before cut point are summarized
3. Summary replaces old messages as "context handoff"
4. Previous compaction summaries chain into new ones

Configuration (~/.pi/agent/settings.json):

{
  "compaction": {
    "enabled": true,
    "reserveTokens": 16384,
    "keepRecentTokens": 20000
  }
}

Note: Compaction is lossy. The agent loses full conversation access afterward. Size tasks to avoid context limits when possible. For critical context, ask the agent to write a summary to a file, then start a new session with that file. The full session history is preserved in the JSONL file; use /branch to revisit any previous point.

Branching

Use /branch to explore alternative conversation paths:

1. Opens selector showing all your user messages
2. Select a message to branch from
3. Creates new session with history up to that point
4. Selected message placed in editor for modification

---

Configuration

Project Context Files

Pi loads AGENTS.md (or CLAUDE.md) files at startup in this order:

1. Global: ~/.pi/agent/AGENTS.md
2. Parent directories: Walking up from current directory
3. Current directory: ./AGENTS.md

Use these for:

• Project instructions and guidelines
• Common commands and workflows
• Architecture documentation
• Coding conventions
• Testing instructions

# Common Commands
- npm run build: Build the project
- npm test: Run tests

# Code Style
- Use TypeScript strict mode
- Prefer async/await over promises

Custom Models and Providers

Add custom models (Ollama, vLLM, LM Studio, etc.) via ~/.pi/agent/models.json:

{
  "providers": {
    "ollama": {
      "baseUrl": "http://localhost:11434/v1",
      "apiKey": "OLLAMA_API_KEY",
      "api": "openai-completions",
      "models": [
        {
          "id": "llama-3.1-8b",
          "name": "Llama 3.1 8B (Local)",
          "reasoning": false,
          "input": ["text"],
          "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0},
          "contextWindow": 128000,
          "maxTokens": 32000
        }
      ]
    }
  }
}

Supported APIs: openai-completions, openai-responses, anthropic-messages, google-generative-ai

API key resolution: The apiKey field is checked as environment variable name first, then used as literal value.

API override: Set api at provider level (default for all models) or model level (override per model).

Custom headers:

{
  "providers": {
    "custom-proxy": {
      "baseUrl": "https://proxy.example.com/v1",
      "apiKey": "YOUR_API_KEY",
      "api": "anthropic-messages",
      "headers": {
        "User-Agent": "Mozilla/5.0 ...",
        "X-Custom-Auth": "token"
      },
      "models": [...]
    }
  }
}

Authorization header: Set authHeader: true to add Authorization: Bearer <apiKey> automatically.

OpenAI compatibility (compat field):

Field	Description
supportsStore	Whether provider supports store field
supportsDeveloperRole	Use developer vs system role
supportsReasoningEffort	Support for reasoning_effort parameter
maxTokensField	Use max_completion_tokens or max_tokens

Live reload: The file reloads each time you open /model. Edit during session; no restart needed.

Model selection priority:

1. CLI args (--provider, --model)
2. First from --models scope (new sessions only)
3. Restored from session (--continue, --resume)
4. Saved default from settings
5. First available model with valid API key

Themes

Built-in themes: dark (default), light. Auto-detected on first run.

/theme  # Interactive selector

Custom themes: Create ~/.pi/agent/themes/*.json. Custom themes support live reload.

mkdir -p ~/.pi/agent/themes
cp $(npm root -g)/@mariozechner/pi-coding-agent/dist/theme/dark.json ~/.pi/agent/themes/my-theme.json

Select with /theme, then edit the file. Changes apply on save.

See Theme Documentation for all 44 color tokens.

VS Code terminal fix: Set terminal.integrated.minimumContrastRatio to 1 for accurate colors.

Custom Slash Commands

Define reusable prompts as Markdown files:

Locations:

• Global: ~/.pi/agent/commands/*.md
• Project: .pi/commands/*.md

Format:

---
description: Review staged git changes
---
Review the staged changes (`git diff --cached`). Focus on:
- Bugs and logic errors
- Security issues
- Error handling gaps

Filename (without .md) becomes the command name. Description shown in autocomplete.

Arguments:

---
description: Create a component
---
Create a React component named $1 with features: $@

Usage: /component Button "onClick handler" "disabled support"

• $1 = Button
• $@ = all arguments joined

Namespacing: Subdirectories create prefixes. .pi/commands/frontend/component.md → /component (project:frontend)

Skills

Skills are instruction files loaded on-demand when tasks match their descriptions. Compatible with Claude Code and Codex CLI skill formats.

Skill locations:

• Pi user: ~/.pi/agent/skills/**/SKILL.md (recursive, colon-separated names)
• Pi project: .pi/skills/**/SKILL.md (recursive, colon-separated names)
• Claude Code user: ~/.claude/skills/*/SKILL.md (one level)
• Claude Code project: .claude/skills/*/SKILL.md (one level)
• Codex CLI: ~/.codex/skills/**/SKILL.md (recursive)

Later locations win on name collisions (Pi skills override Claude/Codex).

Pi skills in subdirectories use colon-separated names: ~/.pi/agent/skills/db/migrate/SKILL.md → db:migrate

Format:

---
description: Extract text and tables from PDF files
---

# PDF Processing

Use `pdftotext` for plain text extraction.
For tables, use `tabula-py`.

Helper scripts: {baseDir}/scripts/

• description: Required. Shown in system prompt for agent to decide when to load.
• name: Optional. Overrides directory name.
• {baseDir}: Placeholder for the skill's directory. Agent substitutes it when following instructions.

How it works:

Skills are listed in the system prompt with descriptions:

<available_skills>
- pdf-extract: Extract text and tables from PDF files
  File: ~/.pi/agent/skills/pdf-extract/SKILL.md
  Base directory: ~/.pi/agent/skills/pdf-extract
</available_skills>

Agent uses read tool to load full instructions when needed.

Disable skills:

CLI: pi --no-skills

Settings (~/.pi/agent/settings.json):

{
  "skills": {
    "enabled": false
  }
}

See docs/skills.md for details.

Hooks

Hooks are TypeScript modules that extend pi's behavior by subscribing to lifecycle events. Use them to:

• Block dangerous commands (permission gates for rm -rf, sudo, etc.)
• Checkpoint code state (git stash at each turn, restore on /branch)
• Protect paths (block writes to .env, node_modules/, etc.)
• Modify tool output (filter or transform results before the LLM sees them)
• Inject messages from external sources (file watchers, webhooks, CI systems)

Hook locations:

• Global: ~/.pi/agent/hooks/*.ts
• Project: .pi/hooks/*.ts
• CLI: --hook <path> (for debugging)

Quick example (permission gate):

import type { HookAPI } from "@mariozechner/pi-coding-agent/hooks";

export default function (pi: HookAPI) {
  pi.on("tool_call", async (event, ctx) => {
    if (event.toolName === "bash" && /sudo/.test(event.input.command as string)) {
      const ok = await ctx.ui.confirm("Allow sudo?", event.input.command as string);
      if (!ok) return { block: true, reason: "Blocked by user" };
    }
    return undefined;
  });
}

Sending messages from hooks:

Use pi.send(text, attachments?) to inject messages into the session. If the agent is streaming, the message is queued; otherwise a new agent loop starts immediately.

import * as fs from "node:fs";
import type { HookAPI } from "@mariozechner/pi-coding-agent/hooks";

export default function (pi: HookAPI) {
  pi.on("session_start", async () => {
    fs.watch("/tmp/trigger.txt", () => {
      const content = fs.readFileSync("/tmp/trigger.txt", "utf-8").trim();
      if (content) pi.send(content);
    });
  });
}

See Hooks Documentation for full API reference.

Settings File

~/.pi/agent/settings.json stores persistent preferences:

{
  "theme": "dark",
  "shellPath": "C:\\path\\to\\bash.exe",
  "queueMode": "one-at-a-time",
  "compaction": {
    "enabled": true,
    "reserveTokens": 16384,
    "keepRecentTokens": 20000
  },
  "skills": {
    "enabled": true
  },
  "retry": {
    "enabled": true,
    "maxRetries": 3,
    "baseDelayMs": 2000
  },
  "terminal": {
    "showImages": true
  }
}

Retry settings:

• enabled: Auto-retry on transient errors (overloaded, rate limit, 5xx). Default: true
• maxRetries: Maximum retry attempts. Default: 3
• baseDelayMs: Base delay for exponential backoff (2s, 4s, 8s). Default: 2000

Terminal settings:

• showImages: Render images inline in supported terminals. Default: true

---

CLI Reference

pi [options] [@files...] [messages...]

Options

Option	Description
--provider <name>	Provider: anthropic, openai, google, mistral, xai, groq, cerebras, openrouter, zai, or custom
--model <id>	Model ID
--api-key <key>	API key (overrides environment)
--system-prompt <text|file>	Custom system prompt (text or file path)
--append-system-prompt <text|file>	Append to system prompt
--mode <mode>	Output mode: text, json, rpc (implies --print)
--print, -p	Non-interactive: process prompt and exit
--no-session	Don't save session
--session <path>	Use specific session file
--continue, -c	Continue most recent session
--resume, -r	Select session to resume
--models <patterns>	Comma-separated patterns for Ctrl+P cycling (e.g., sonnet:high,haiku:low)
--tools <tools>	Comma-separated tool list (default: read,bash,edit,write)
--thinking <level>	Thinking level: off, minimal, low, medium, high
--hook <path>	Load a hook file (can be used multiple times)
--no-skills	Disable skills discovery and loading
--export <file> [output]	Export session to HTML
--help, -h	Show help
--version, -v	Show version

File Arguments

Include files with @ prefix:

pi @prompt.md "Answer this"
pi @screenshot.png "What's in this image?"
pi @requirements.md @design.png "Implement this"

Text files wrapped in <file name="path">content</file>. Images attached as base64.

Examples

# Interactive mode
pi

# Interactive with initial prompt
pi "List all .ts files in src/"

# Non-interactive
pi -p "List all .ts files in src/"

# With files
pi -p @code.ts "Review this code"

# JSON event stream
pi --mode json "List files"

# RPC mode (headless)
pi --mode rpc --no-session

# Continue session
pi -c "What did we discuss?"

# Specific model
pi --provider openai --model gpt-4o "Help me refactor"

# Model cycling with thinking levels
pi --models sonnet:high,haiku:low

# Read-only mode
pi --tools read,grep,find,ls -p "Review the architecture"

# Export session
pi --export session.jsonl output.html

---

Tools

Default Tools

Tool	Description
read	Read file contents. Images sent as attachments. Text: first 2000 lines, lines truncated at 2000 chars. Use offset/limit for large files.
write	Write/overwrite file. Creates parent directories.
edit	Replace exact text in file. Must match exactly including whitespace. Fails if text appears multiple times or not found.
bash	Execute command. Returns stdout/stderr. Optional timeout parameter.

Read-Only Tools

Available via --tools flag:

Tool	Description
grep	Search file contents (regex or literal). Respects .gitignore.
find	Search for files by glob pattern. Respects .gitignore.
ls	List directory contents. Includes dotfiles.

Example: --tools read,grep,find,ls for code review without modification.

Custom Tools

Pi relies on CLI tools invoked via bash rather than MCP. Create a tool with a README:

~/agent-tools/screenshot/README.md:

# Screenshot Tool
Takes a screenshot of your main display.

## Usage
```bash
screenshot.sh

Returns the path to the saved PNG.

`~/agent-tools/screenshot/screenshot.sh`:
```bash
#!/bin/bash
screencapture -x /tmp/screenshot-$(date +%s).png
ls -t /tmp/screenshot-*.png | head -1

Usage: "Read ~/agent-tools/screenshot/README.md and take a screenshot"

Reference tool READMEs in AGENTS.md to make them automatically available.

---

Programmatic Usage

RPC Mode

For embedding pi in other applications:

pi --mode rpc --no-session

Send JSON commands on stdin:

{"type":"prompt","message":"List all .ts files"}
{"type":"abort"}

See RPC documentation for full protocol.

Node.js/TypeScript: Consider using AgentSession directly from @mariozechner/pi-coding-agent instead of subprocess. See src/core/agent-session.ts and src/modes/rpc/rpc-client.ts.

HTML Export

pi --export session.jsonl              # Auto-generated filename
pi --export session.jsonl output.html  # Custom filename

Works with both session files and streaming event logs from --mode json.

---

Philosophy

Pi is opinionated about what it won't do. These are intentional design decisions.

No MCP

Pi does not support MCP (Model Context Protocol). Instead, it relies on four core tools (read, write, edit, bash) and assumes the agent can invoke CLI tools or write them as needed.

CLI tools are simpler: any executable with a README works. No protocol overhead, no server management. The agent reads the README and uses bash.

See: What if you don't need MCP?

No Sub-Agents

If the agent needs to delegate, it can spawn pi via bash or write a custom tool. Built-in sub-agents transfer context poorly; information gets lost or misrepresented. For parallel work, run multiple pi sessions in different terminals.

No Built-in To-Dos

To-do lists confuse models more than they help. For task tracking, use a file:

# TODO.md
- [x] Implement authentication
- [ ] Write API docs

No Planning Mode

Tell the agent to think through problems without modifying files. For persistent plans, write to a file:

# PLAN.md
## Goal
Refactor auth to support OAuth
## Current Step
Working on authorization endpoints

No Permission System (YOLO Mode)

Pi runs with full filesystem access and no permission prompts. Why:

• Permission systems add friction while being easily circumvented
• Pre-checking for "dangerous" patterns causes latency and false positives

Risks:

• Can read, write, delete anything with your user privileges
• Prompt injection via files or command output can influence behavior

Mitigations:

• Run in a container if uncomfortable
• Don't use on systems with sensitive data you can't afford to lose

No Background Bash

Use tmux or similar. Bonus: you can watch the agent interact with CLIs and intervene if needed.

---

Development

Forking / Rebranding

Configure via package.json:

{
  "piConfig": {
    "name": "pi",
    "configDir": ".pi"
  }
}

Change name, configDir, and bin field for your fork. Affects CLI banner, config paths, and environment variable names.

Path Resolution

Three execution modes: npm install, standalone binary, tsx from source.

Always use src/paths.ts for package assets:

import { getPackageDir, getThemeDir } from "./paths.js";

Never use __dirname directly for package assets.

Debug Command

/debug (hidden) writes rendered lines with ANSI codes to ~/.pi/agent/pi-debug.log for TUI debugging.

For architecture and contribution guidelines, see DEVELOPMENT.md.

---

License

MIT

See Also

• @mariozechner/pi-ai: Core LLM toolkit
• @mariozechner/pi-agent: Agent framework