Overview

SHIPIT Agent is a standalone Python agent library focused on a clean runtime:

• bring your own LLM — or use any of seven built-in provider adapters
• attach Python tools, remote MCP servers, or connector-style third-party tools (Gmail, Drive, Slack, Linear, Notion, Jira, Confluence)
• attach packaged or custom skills to steer agent behavior and reusable workflows
• iterate tool-using agents with configurable retry and router policies
• stream structured events (including reasoning / thinking blocks) as they happen
• inspect every step: reasoning, tool arguments, tool outputs, retries, final answer
• compose reusable agent profiles with system prompts and tool selections locked in
• keep clean boundaries between runtime, tools, MCP, policies, and profiles

Built for developers who want the agent loop observable, interchangeable, and out of the way.

Install

pip install shipit-agent

With optional extras:

pip install 'shipit-agent[openai]'         # OpenAI SDK
pip install 'shipit-agent[anthropic]'      # Anthropic SDK (native thinking blocks)
pip install 'shipit-agent[litellm]'        # LiteLLM (Bedrock, Gemini, Groq, Together, …)
pip install 'shipit-agent[playwright]'     # In-process browser for open_url and web_search
pip install 'shipit-agent[all]'            # Everything

30-second example

from shipit_agent import Agent
from shipit_agent.llms import OpenAIChatLLM

agent = Agent.with_builtins(llm=OpenAIChatLLM(model="gpt-4o-mini"))

for event in agent.stream("Search the web for today's Bitcoin price in USD."):
    print(event.type, event.message)

Emits events like:

run_started           Agent run started
step_started          LLM completion started
reasoning_started     🧠 Model reasoning started
reasoning_completed   🧠 Model reasoning completed
tool_called           Tool called: web_search
tool_completed        Tool completed: web_search
run_completed         Agent run completed

Why SHIPIT Agent

Next steps

• Install and run the quick start — get an agent running in five minutes
• Explore streaming events — understand the 14 event types and what they carry
• Reasoning and thinking steps — render a live "Thinking" panel in your UI
• Create a custom tool — build a new tool from scratch
• Use skills — packaged skills, custom skills, Agent, and DeepAgent workflows
• MCP integration — attach remote MCP servers to extend capabilities
• Parallel tool execution — speed up multi-tool turns
• Hooks & middleware — add cost tracking, logging, and guardrails
• Async runtime — use with FastAPI and async Python
• Context window management — track tokens and manage context limits
• Error recovery — graceful failure handling and retry policies

Try it now — runnable examples

The repo ships with 7 numbered, copy-pasteable examples covering every major feature. Pick one and run it in 30 seconds.

See the full examples README →

Provider compatibility matrix

Tip: if you want a "Thinking" panel UI without paying for o1/Claude, AWS Bedrock's openai.gpt-oss-120b-1:0 is the cheapest reasoning-capable model in the matrix and ships with Agent.with_builtins(llm=BedrockChatLLM()) out of the box.

What you get vs. what you don't

This is a library, not a framework. The runtime is small enough to read in one sitting (shipit_agent/runtime.py is under 400 lines). Bring your own LLM, tools, and storage; the runtime composes them and gets out of the way.