Skip to content

pi.md

Latest commit

 

History

History
115 lines (80 loc) · 4.21 KB

pi.md

File metadata and controls

115 lines (80 loc) · 4.21 KB

Pi Coding Agent

OpenAB supports the Pi coding agent via the pi-acp adapter — a Node.js bridge that translates ACP JSON-RPC into Pi CLI invocations.

Advantages Over Other Native Coding Agents

Pi is a native coding agent that supports subscription-based authentication (like Codex, Cloud Code, and GitHub Copilot). Key advantages:

No Auth Proxy Required

Pi natively supports Anthropic (Claude Pro/Max) and ChatGPT Plus/Pro subscriptions via OAuth. Unlike agents that require an openab-auth-proxy sidecar for subscription forwarding, Pi handles subscription auth directly — reducing deployment complexity and eliminating a moving part.

Agent Subscription Auth Auth Proxy Needed?
Pi Native OAuth (pi /login) ❌ No
Codex Native device flow ❌ No
GitHub Copilot Native device flow ❌ No
Claude Code Native OAuth ❌ No
Kiro Native OAuth ❌ No

Minimal Tool Surface (Maximum Context Window)

Pi exposes only 4 core tools: read, write, edit, bash. Combined with a tiny system prompt, this drastically reduces prompt overhead and maximizes the available context window for actual project source files.

Agent Tool Count System Prompt Size
Pi 4 Minimal
Claude Code 10+ Large
Codex 8+ Medium
Copilot 10+ Large

Multi-Model Support

Pi is model-agnostic and supports 15+ LLM providers. Developers can switch models mid-session without restarting the agent or changing configuration.

Supported providers include:

  • Anthropic (Claude) — via subscription or API key
  • OpenAI (GPT/Codex) — via subscription or API key
  • Google (Gemini) — via API key
  • Any OpenAI-compatible endpoint

Branching Session Trees

Pi saves session history as trees, enabling clean branching of code exploration. This allows developers to explore multiple approaches from a single decision point without losing context.

Configuration

[agent]
# command = "pi-acp"  # optional — defaults from OPENAB_AGENT_COMMAND
# working_dir = "/home/node"  # optional — defaults to $HOME

Docker

docker build -f Dockerfile.pi -t openab-pi:latest .

Helm

agents:
  pi:
    discord:
      enabled: true
      allowedChannels:
        - "YOUR_CHANNEL_ID"
    command: pi-acp
    workingDir: /home/node
    image: "ghcr.io/openabdev/openab-pi:latest"

Authentication

kubectl exec -it deployment/openab-pi -- pi
# Once inside the interactive interface, type /login to authenticate

Supported authentication methods:

Provider Auth Method Subscription
Anthropic (Claude Pro/Max) OAuth via pi /login Claude subscription
ChatGPT Plus/Pro OAuth via pi /login ChatGPT subscription
Any API key provider env = { OPENAI_API_KEY = "..." } Pay-per-token

Local OpenAI-Compatible Vision Models

OpenAB can pass inbound image attachments to Pi as ACP image content blocks, but Pi must also select a model declared as image-capable. For custom OpenAI-compatible providers, add input: ["text", "image"] to the model entry in ~/.pi/agent/models.json.

See Local OpenAI-Compatible Vision Models for the llama-server setup, models.json example, and local vision pitfalls.

Steering Files

Pi reads steering files in this order:

  1. .pi/SYSTEM.md — replaces the default system prompt entirely
  2. .pi/APPEND_SYSTEM.md — appends to the default system prompt
  3. AGENTS.md — loaded hierarchically (project root → global) for context injection

Place your steering instructions in /home/node/AGENTS.md or /home/node/.pi/APPEND_SYSTEM.md.

Persisted Paths (PVC)

Path Contents
/home/node/.pi/ Pi configuration and auth tokens
/home/node/.pi/sessions/ Session history trees

Limitations

  • No streaming: pi-acp returns the full response at once; streamed output is sent as a single agent_message_chunk notification.
  • Cancel is best-effort: Pi CLI runs to completion; session/cancel may not interrupt mid-generation.