co-mono/packages/coding-agent

pi

A radically simple and opinionated coding agent with multi-model support (including mid-session switching), a simple yet powerful CLI for headless coding tasks, and many creature comforts you might be used to from other coding agents.

Works on Linux, macOS, and Windows (barely tested, needs Git Bash running in the "modern" Windows Terminal).

Table of Contents

• Installation
• Quick Start
• API Keys
• OAuth Authentication (Optional)
• Custom Models and Providers
• Slash Commands
• Editor Features
• Project Context Files
• Image Support
• Session Management
• CLI Options
• Tools
• Usage
• Security (YOLO by default)
• Sub-Agents
• To-Dos
• Planning
• Background Bash
• Planned Features
• License
• See Also

Installation

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

Quick Start

# Set your API key (see API Keys section)
export ANTHROPIC_API_KEY=sk-ant-...

# Start the interactive CLI
pi

Once in the CLI, you can chat with the AI:

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

The agent will use its tools to read, write, and edit files as needed, and execute commands via Bash.

API Keys

The CLI supports multiple LLM providers. Set the appropriate environment variable for your chosen provider:

# Anthropic (Claude)
export ANTHROPIC_API_KEY=sk-ant-...
# Or use OAuth token (retrieved via: claude setup-token)
export ANTHROPIC_OAUTH_TOKEN=...

# OpenAI (GPT)
export OPENAI_API_KEY=sk-...

# Google (Gemini)
export GEMINI_API_KEY=...

# Groq
export GROQ_API_KEY=gsk_...

# Cerebras
export CEREBRAS_API_KEY=csk-...

# xAI (Grok)
export XAI_API_KEY=xai-...

# OpenRouter
export OPENROUTER_API_KEY=sk-or-...

# ZAI
export ZAI_API_KEY=...

If no API key is set, the CLI will prompt you to configure one on first run.

Note: The /model command only shows models for which API keys are configured in your environment. If you don't see a model you expect, check that you've set the corresponding environment variable.

OAuth Authentication (Optional)

If you have a Claude Pro/Max subscription, you can use OAuth instead of API keys:

pi
# In the interactive session:
/login
# Select "Anthropic (Claude Pro/Max)"
# Authorize in browser
# Paste authorization code

This gives you:

• Free access to Claude models (included in your subscription)
• No need to manage API keys
• Automatic token refresh

To logout:

/logout

Note: OAuth tokens are stored in ~/.pi/agent/oauth.json with restricted permissions (0600).

Custom Models and Providers

You can add custom models and providers (like Ollama, vLLM, LM Studio, or any custom API endpoint) via ~/.pi/agent/models.json. Supports OpenAI-compatible APIs (openai-completions, openai-responses), Anthropic Messages API (anthropic-messages), and Google Generative AI API (google-generative-ai). This file is loaded fresh every time you open the /model selector, allowing live updates without restarting.

Configuration File Structure

{
  "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
        }
      ]
    },
    "vllm": {
      "baseUrl": "http://your-server:8000/v1",
      "apiKey": "VLLM_API_KEY",
      "api": "openai-completions",
      "models": [
        {
          "id": "custom-model",
          "name": "Custom Fine-tuned Model",
          "reasoning": false,
          "input": ["text", "image"],
          "cost": {"input": 0.5, "output": 1.0, "cacheRead": 0, "cacheWrite": 0},
          "contextWindow": 32768,
          "maxTokens": 8192
        }
      ]
    },
    "mixed-api-provider": {
      "baseUrl": "https://api.example.com/v1",
      "apiKey": "CUSTOM_API_KEY",
      "api": "openai-completions",
      "models": [
        {
          "id": "legacy-model",
          "name": "Legacy Model",
          "reasoning": false,
          "input": ["text"],
          "cost": {"input": 1.0, "output": 2.0, "cacheRead": 0, "cacheWrite": 0},
          "contextWindow": 8192,
          "maxTokens": 4096
        },
        {
          "id": "new-model",
          "name": "New Model",
          "api": "openai-responses",
          "reasoning": true,
          "input": ["text", "image"],
          "cost": {"input": 0.5, "output": 1.0, "cacheRead": 0.1, "cacheWrite": 0.2},
          "contextWindow": 128000,
          "maxTokens": 32000
        }
      ]
    }
  }
}

API Key Resolution

The apiKey field can be either an environment variable name or a literal API key:

1. First, pi checks if an environment variable with that name exists
2. If found, uses the environment variable's value
3. Otherwise, treats it as a literal API key

Examples:

• "apiKey": "OLLAMA_API_KEY" → checks $OLLAMA_API_KEY, then treats as literal "OLLAMA_API_KEY"
• "apiKey": "sk-1234..." → checks $sk-1234... (unlikely to exist), then uses literal value

This allows both secure env var usage and literal keys for local servers.

API Override

• Provider-level api: Sets the default API for all models in that provider
• Model-level api: Overrides the provider default for specific models
• Supported APIs: openai-completions, openai-responses, anthropic-messages, google-generative-ai

This is useful when a provider supports multiple API standards through the same base URL.

Custom Headers

You can add custom HTTP headers to bypass Cloudflare bot detection, add authentication tokens, or meet other proxy requirements:

{
  "providers": {
    "custom-proxy": {
      "baseUrl": "https://proxy.example.com/v1",
      "apiKey": "YOUR_API_KEY",
      "api": "anthropic-messages",
      "headers": {
        "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36",
        "X-Custom-Auth": "bearer-token-here"
      },
      "models": [
        {
          "id": "claude-sonnet-4",
          "name": "Claude Sonnet 4 (Proxied)",
          "reasoning": true,
          "input": ["text", "image"],
          "cost": {"input": 3, "output": 15, "cacheRead": 0.3, "cacheWrite": 3.75},
          "contextWindow": 200000,
          "maxTokens": 8192,
          "headers": {
            "X-Model-Specific-Header": "value"
          }
        }
      ]
    }
  }
}

• Provider-level headers: Applied to all requests for models in that provider
• Model-level headers: Additional headers for specific models (merged with provider headers)
• Model headers override provider headers when keys conflict

Model Selection Priority

When starting pi, models are selected in this order:

1. CLI args: --provider and --model flags
2. Restored from session: If using --continue or --resume
3. Saved default: From ~/.pi/agent/settings.json (set when you select a model with /model)
4. First available: First model with a valid API key
5. None: Allowed in interactive mode (shows error on message submission)

Provider Defaults

When multiple providers are available, pi prefers sensible defaults before falling back to "first available":

Provider	Default Model
anthropic	claude-sonnet-4-5
openai	gpt-5.1-codex
google	gemini-2.5-pro
openrouter	openai/gpt-5.1-codex
xai	grok-4-fast-non-reasoning
groq	openai/gpt-oss-120b
cerebras	zai-glm-4.6
zai	glm-4.6

Live Reload & Errors

The models.json file is reloaded every time you open the /model selector. This means:

• Edit models.json during a session
• Or have the agent write/update it for you
• Use /model to see changes immediately
• No restart needed!

If the file contains errors (JSON syntax, schema violations, missing fields), the selector shows the exact validation error and file path in red so you can fix it immediately.

Example: Adding Ollama Models

See the configuration structure above. Create ~/.pi/agent/models.json with your Ollama setup, then use /model to select your local models. The agent can also help you write this file if you point it to this README.

Slash Commands

The CLI supports several commands to control its behavior:

/model

Switch models mid-session. Opens an interactive selector where you can type to search (by provider or model name), use arrow keys to navigate, Enter to select, or Escape to cancel.

The selector only displays models for which API keys are configured in your environment (see API Keys section).

/thinking

Adjust thinking/reasoning level for supported models (Claude Sonnet 4, GPT-5, Gemini 2.5). Opens an interactive selector where you can use arrow keys to navigate, Enter to select, or Escape to cancel.

/queue

Select message queue mode. Opens an interactive selector where you can choose between:

• one-at-a-time (default): Process queued messages one by one. When you submit messages while the agent is processing, they're queued and sent individually after each agent response completes.
• all: Process all queued messages at once. All queued messages are injected into the context together before the next agent response.

The queue mode setting is saved and persists across sessions.

/export [filename]

Export the current session to a self-contained HTML file:

/export                          # Auto-generates filename
/export my-session.html          # Custom filename

The HTML file includes the full conversation with syntax highlighting and is viewable in any browser.

/session

Show session information and statistics:

/session

Displays:

• Session file path and ID
• Message counts (user, assistant, total)
• Token usage (input, output, cache read/write, total)
• Total cost (if available)

/changelog

Display the full changelog with all version history (newest last):

/changelog

/branch

Create a new conversation branch from a previous message. Opens an interactive selector showing all your user messages in chronological order. Select a message to:

1. Create a new session with all messages before the selected one
2. Place the selected message in the editor for modification or resubmission

This allows you to explore alternative conversation paths without losing your current session.

/branch

/login

Login with OAuth to use subscription-based models (Claude Pro/Max):

/login

Opens an interactive selector to choose provider, then guides you through the OAuth flow in your browser.

/logout

Logout from OAuth providers:

/logout

Shows a list of logged-in providers to logout from.

Editor Features

The interactive input editor includes several productivity features:

Path Completion

Press Tab to autocomplete file and directory paths:

• Works with relative paths: ./src/ + Tab → complete files in src/
• Works with parent directories: ../../ + Tab → navigate up and complete
• Works with home directory: ~/Des + Tab → ~/Desktop/
• Use Up/Down arrows to navigate completion suggestions
• Press Enter to select a completion
• Shows matching files and directories as you type

File Drag & Drop

Drag files from your OS file explorer (Finder on macOS, Explorer on Windows) directly onto the terminal. The file path will be automatically inserted into the editor. Works great with screenshots from macOS screenshot tool.

Multi-line Paste

Paste multiple lines of text (e.g., code snippets, logs) and they'll be automatically coalesced into a compact [paste #123 <N> lines] reference in the editor. The full content is still sent to the model.

Message Queuing

You can submit multiple messages while the agent is processing without waiting for responses. Messages are queued and processed based on your queue mode setting:

One-at-a-time mode (default):

• Each queued message is processed sequentially with its own response
• Example: Queue "task 1", "task 2", "task 3" → agent completes task 1 → processes task 2 → completes task 2 → processes task 3
• Recommended for most use cases

All mode:

• All queued messages are sent to the model at once in a single context
• Example: Queue "task 1", "task 2", "task 3" → agent receives all three together → responds considering all tasks
• Useful when tasks should be considered together

Visual feedback:

• Queued messages appear below the chat with "Queued: "
• Messages disappear from the queue as they're processed

Abort and restore:

• Press Escape while streaming to abort the current operation
• All queued messages (plus any text in the editor) are restored to the editor
• Allows you to modify or remove queued messages before resubmitting

Change queue mode with /queue command. Setting is saved in ~/.pi/agent/settings.json.

Keyboard Shortcuts

• Ctrl+W: Delete word backwards (stops at whitespace or punctuation)
• Option+Backspace (Ghostty): Delete word backwards (same as Ctrl+W)
• Ctrl+U: Delete to start of line (at line start: merge with previous line)
• Cmd+Backspace (Ghostty): Delete to start of line (same as Ctrl+U)
• Ctrl+K: Delete to end of line (at line end: merge with next line)
• Ctrl+C: Clear editor (first press) / Exit pi (second press)
• Tab: Path completion
• Shift+Tab: Cycle thinking level (for reasoning-capable models)
• Ctrl+P: Cycle models (use --models to scope)
• Ctrl+O: Toggle tool output expansion (collapsed ↔ full output)
• Enter: Send message
• Shift+Enter: Insert new line (multi-line input)
• Backspace: Delete character backwards
• Delete (or Fn+Backspace): Delete character forwards
• Arrow keys: Move cursor (Up/Down/Left/Right)
• Ctrl+A / Home / Cmd+Left (macOS): Jump to start of line
• Ctrl+E / End / Cmd+Right (macOS): Jump to end of line
• Escape: Cancel autocomplete (when autocomplete is active)

Project Context Files

The agent automatically loads context from AGENTS.md or CLAUDE.md files at the start of new sessions (not when continuing/resuming). These files are loaded in hierarchical order to support both global preferences and monorepo structures.

File Locations

Context files are loaded in this order:

1. Global context: ~/.pi/agent/AGENTS.md or CLAUDE.md

   • Applies to all your coding sessions
   • Great for personal coding preferences and workflows

2. Parent directories (top-most first down to current directory)

   • Walks up from current directory to filesystem root
   • Each directory can have its own AGENTS.md or CLAUDE.md
   • Perfect for monorepos with shared context at higher levels

3. Current directory: Your project's AGENTS.md or CLAUDE.md

   • Most specific context, loaded last
   • Overwrites or extends parent/global context

File preference: In each directory, AGENTS.md is preferred over CLAUDE.md if both exist.

What to Include

Context files are useful for:

• Project-specific instructions and guidelines
• Common bash commands and workflows
• Architecture documentation
• Coding conventions and style guides
• Dependencies and setup information
• Testing instructions
• Repository etiquette (branch naming, merge vs. rebase, etc.)

Example

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

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

# Workflow
- Always run tests before committing
- Update CHANGELOG.md for user-facing changes

All context files are automatically included in the system prompt at session start, along with the current date/time and working directory. This ensures the AI has complete project context from the very first message.

Image Support

Send images to vision-capable models by providing file paths:

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

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

The image will be automatically encoded and sent with your message. JPEG and PNG are supported across all vision models. Other formats may only be supported by some models.

Session Management

Sessions are automatically saved in ~/.pi/agent/sessions/ organized by working directory. Each session is stored as a JSONL file with a unique timestamp-based ID.

To continue the most recent session:

pi --continue
# or
pi -c

To browse and select from past sessions:

pi --resume
# or
pi -r

This opens an interactive session selector where you can:

• Type to search through session messages
• Use arrow keys to navigate the list
• Press Enter to resume a session
• Press Escape to cancel

Sessions include all conversation messages, tool calls and results, model switches, and thinking level changes.

To run without saving a session (ephemeral mode):

pi --no-session

To use a specific session file instead of auto-generating one:

pi --session /path/to/my-session.jsonl

CLI Options

pi [options] [messages...]

Options

--provider Provider name. Available: anthropic, openai, google, xai, groq, cerebras, openrouter, zai, plus any custom providers defined in ~/.pi/agent/models.json.

--model Model ID. If not specified, uses: (1) saved default from settings, (2) first available model with valid API key, or (3) none (interactive mode only).

--api-key API key (overrides environment variables)

--system-prompt <text|file> Custom system prompt. Can be:

• Inline text: --system-prompt "You are a helpful assistant"
• File path: --system-prompt ./my-prompt.txt

If the argument is a valid file path, the file contents will be used as the system prompt. Otherwise, the text is used directly. Project context files and datetime are automatically appended.

--mode Output mode for non-interactive usage. Options:

• text (default): Output only the final assistant message text
• json: Stream all agent events as JSON (one event per line). Events are emitted by @mariozechner/pi-agent and include message updates, tool executions, and completions
• rpc: JSON mode plus stdin listener for headless operation. Send JSON commands on stdin: {"type":"prompt","message":"..."} or {"type":"abort"}. See test/rpc-example.ts for a complete example

--no-session Don't save session (ephemeral mode)

--session Use specific session file path instead of auto-generating one

--continue, -c Continue the most recent session

--resume, -r Select a session to resume (opens interactive selector)

--models Comma-separated model patterns for quick cycling with Ctrl+P. Patterns match against model IDs and names (case-insensitive). When multiple versions exist, prefers aliases over dated versions (e.g., claude-sonnet-4-5 over claude-sonnet-4-5-20250929). Without this flag, Ctrl+P cycles through all available models.

Examples:

• --models claude-sonnet,gpt-4o - Scope to Claude Sonnet and GPT-4o
• --models sonnet,haiku - Match any model containing "sonnet" or "haiku"
• --models gemini - All Gemini models

--help, -h Show help message

Examples

# Start interactive mode
pi

# Single message mode (text output)
pi "List all .ts files in src/"

# JSON mode - stream all agent events
pi --mode json "List all .ts files in src/"

# RPC mode - headless operation (see test/rpc-example.ts)
pi --mode rpc --no-session
# Then send JSON on stdin:
# {"type":"prompt","message":"List all .ts files"}
# {"type":"abort"}

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

# Use different model
pi --provider openai --model gpt-4o "Help me refactor this code"

# Limit model cycling to specific models
pi --models claude-sonnet,claude-haiku,gpt-4o
# Now Ctrl+P cycles only through those models

Tools

Built-in Tools

The agent has access to four core tools for working with your codebase:

read Read file contents. Supports text files and images (jpg, png, gif, webp). Images are sent as attachments. For text files, defaults to first 2000 lines. Use offset/limit parameters for large files. Lines longer than 2000 characters are truncated.

write Write content to a file. Creates the file if it doesn't exist, overwrites if it does. Automatically creates parent directories.

edit Edit a file by replacing exact text. The oldText must match exactly (including whitespace). Use this for precise, surgical edits. Returns an error if the text appears multiple times or isn't found.

bash Execute a bash command in the current working directory. Returns stdout and stderr. Optionally accepts a timeout parameter (in seconds) - no default timeout.

MCP & Adding Your Own Tools

pi does and will not support MCP. Instead, it relies on the four built-in tools above and assumes the agent can invoke pre-existing CLI tools or write them on the fly as needed.

Here's the gist:

1. Create a simple CLI tool (any language, any executable)
2. Write a concise README.md describing what it does and how to use it
3. Tell the agent to read that README

Minimal example:

~/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 file.

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

In your session:

You: Read ~/agent-tools/screenshot/README.md and use that tool to take a screenshot

The agent will read the README, understand the tool, and invoke it via bash as needed. If you need a new tool, ask the agent to write it for you.

You can also reference tool READMEs in your AGENTS.md files to make them automatically available:

• Global: ~/.pi/agent/AGENTS.md - available in all sessions
• Project-specific: ./AGENTS.md - available in this project

Real-world example:

The exa-search tools provide web search capabilities via the Exa API. Built by the agent itself in ~2 minutes. Far from perfect, but functional. Just tell your agent: "Read ~/agent-tools/exa-search/README.md and search for X".

For a detailed walkthrough with more examples, and the reasons for and benefits of this decision, see: https://mariozechner.at/posts/2025-11-02-what-if-you-dont-need-mcp/

Security (YOLO by default)

This agent runs in full YOLO mode and assumes you know what you're doing. It has unrestricted access to your filesystem and can execute any command without permission checks or safety rails.

What this means:

• No permission prompts for file operations or commands
• No pre-checking of bash commands for malicious content
• Full filesystem access - can read, write, or delete anything
• Can execute any command with your user privileges

Why:

• Permission systems add massive friction while being easily circumvented
• Pre-checking tools for "dangerous" patterns introduces latency, false positives, and is ineffective

Prompt injection risks:

• By default, pi has no web search or fetch tool
• However, it can use curl or read files from disk
• Both provide ample surface area for prompt injection attacks
• Malicious content in files or command outputs can influence behavior

Mitigations:

• Run pi inside a container if you're uncomfortable with full access
• Use a different tool if you need guardrails
• Don't use pi on systems with sensitive data you can't afford to lose
• Fork pi and add all of the above

This is how I want it to work and I'm not likely to change my stance on this.

Use at your own risk.

Sub-Agents

pi does not and will not support sub-agents as a built-in feature. If the agent needs to delegate work, it can:

1. Spawn another instance of itself via the pi CLI command
2. Write a custom tool with a README.md that describes how to invoke pi for specific tasks

Why no built-in sub-agents:

Context transfer between agents is generally poor. Information gets lost, compressed, or misrepresented when passed through agent boundaries. Direct execution with full context is more effective than delegation with summarized context.

If you need parallel work on independent tasks, manually run multiple pi sessions in different terminal tabs. You're the orchestrator.

To-Dos

pi does not and will not support built-in to-dos. In my experience, to-do lists generally confuse models more than they help.

If you need task tracking, make it stateful by writing to a file:

# TODO.md

- [x] Implement user authentication
- [x] Add database migrations
- [ ] Write API documentation
- [ ] Add rate limiting

The agent can read and update this file as needed. Using checkboxes keeps track of what's done and what remains. Simple, visible, and under your control.

Planning

pi does not and will not have a built-in planning mode. Telling the agent to think through a problem together with you, without modifying files or executing commands, is generally sufficient.

If you need persistent planning across sessions, write it to a file:

# PLAN.md

## Goal
Refactor authentication system to support OAuth

## Approach
1. Research OAuth 2.0 flows
2. Design token storage schema
3. Implement authorization server endpoints
4. Update client-side login flow
5. Add tests

## Current Step
Working on step 3 - authorization endpoints

The agent can read, update, and reference the plan as it works. Unlike ephemeral planning modes that only exist within a session, file-based plans persist and can be versioned with your code.

Background Bash

pi does not and will not implement background bash execution. Instead, tell the agent to use tmux or something like tterminal-cp. Bonus points: you can watch the agent interact with a CLI like a debugger and even intervene if necessary.

Planned Features

Things that might happen eventually:

• Auto-compaction: Currently, watch the context percentage at the bottom. When it approaches 80%, either:
  • Ask the agent to write a summary .md file you can load in a new session
  • Switch to a model with bigger context (e.g., Gemini) using /model and either continue with that model, or let it summarize the session to a .md file to be loaded in a new session
• Better RPC mode docs: It works, you'll figure it out (see test/rpc-example.ts)

License

MIT

See Also

• @mariozechner/pi-ai: Core LLM toolkit with multi-provider support
• @mariozechner/pi-agent: Agent framework with tool execution