OpenAI Agents SDK: GPT & Ollama Backend

The OpenAI Agents SDK backend uses OpenAI’s native agent framework (openai-agents package) with GPT models. It also supports Ollama and OpenAI-compatible providers for local inference.

Overview

  • Status: Beta
  • Provider support: OpenAI, Ollama, OpenAI-compatible, OpenRouter, LiteLLM
  • Built-in tools: code_interpreter, file_search, computer_use
  • MCP support: No
  • Session persistence: SQLite-based via the SDK

Configuration

Terminal window
export POCKETPAW_AGENT_BACKEND="openai_agents"
export POCKETPAW_OPENAI_API_KEY="sk-..."

Backend-Specific Settings

SettingEnv VariableDefaultDescription
openai_agents_providerPOCKETPAW_OPENAI_AGENTS_PROVIDER"" (auto)Provider: openai, ollama, openai_compatible, openrouter, or litellm
openai_agents_modelPOCKETPAW_OPENAI_AGENTS_MODEL"" (auto)Model to use (e.g., gpt-4o, gpt-4o-mini)
openai_agents_max_turnsPOCKETPAW_OPENAI_AGENTS_MAX_TURNS0 (unlimited)Max tool-use turns per query

Using with Ollama

For local inference via Ollama’s OpenAI-compatible API:

Terminal window
export POCKETPAW_AGENT_BACKEND="openai_agents"
export POCKETPAW_OPENAI_AGENTS_PROVIDER="ollama"
export POCKETPAW_OLLAMA_MODEL="qwen2.5:7b"

The backend wraps the Ollama model in OpenAIChatCompletionsModel from the SDK.

Using with OpenRouter

Terminal window
export POCKETPAW_AGENT_BACKEND="openai_agents"
export POCKETPAW_OPENAI_AGENTS_PROVIDER="openrouter"
export POCKETPAW_OPENROUTER_API_KEY="sk-or-v1-..."
export POCKETPAW_OPENROUTER_MODEL="anthropic/claude-sonnet-4-6"

Using with LiteLLM

Proxy mode (recommended when running a LiteLLM proxy):

Terminal window
export POCKETPAW_AGENT_BACKEND="openai_agents"
export POCKETPAW_OPENAI_AGENTS_PROVIDER="litellm"
export POCKETPAW_LITELLM_API_BASE="http://localhost:4000"
export POCKETPAW_LITELLM_MODEL="gpt-4o"

Direct SDK mode (no proxy, LiteLLM SDK calls providers directly):

Terminal window
export POCKETPAW_AGENT_BACKEND="openai_agents"
export POCKETPAW_OPENAI_AGENTS_PROVIDER="litellm"
export POCKETPAW_LITELLM_API_BASE=""
export POCKETPAW_LITELLM_API_KEY="sk-ant-..."
export POCKETPAW_LITELLM_MODEL="anthropic/claude-sonnet-4-6"  # must be LiteLLM-prefixed
Warning

In direct SDK mode (no LITELLM_API_BASE), model names must use LiteLLM’s provider prefix format (e.g., anthropic/claude-sonnet-4-6, openai/gpt-4o). Without the prefix, LiteLLM cannot determine which provider to route to.

Built-in Tools

SDK ToolPolicy Mapping
code_interpretershell
file_searchread_file
computer_useshell

PocketPaw’s custom tools (web_search, image_gen, etc.) are also registered via the tool_bridge.py wrapper.

How It Works

  1. PocketPaw creates an OpenAI Agent with the system prompt and tool definitions
  2. Messages are sent via Runner.run_streamed()
  3. The SDK handles multi-turn tool calling automatically
  4. Responses stream back as AgentEvent objects
  5. Session history is managed via SQLiteSession for multi-turn context

Installation

Terminal window
pip install pocketpaw[openai-agents]

This installs the openai-agents package as an optional dependency.

Ollama: Free Local Models

Run the OpenAI Agents backend with local models via Ollama.

Claude Agent SDK

The recommended backend with native tools and MCP integration.

All Agent Backends

Compare all six backends side by side with capability badges.

Local LLM Agent Guide

Step-by-step guide to running PocketPaw with local models.

Last updated: April 29, 2026
2 min read
Edit this page