An agent framework whose killer feature is restraint

What it does

Stirrup is a lightweight Python framework and starting template for building agents. You get an agent loop, a session() context manager that handles tool lifecycle, logging, and file outputs, and a generic Tool interface with Pydantic-typed parameters. Default tools ship with it — shell execution in an isolated temp directory plus web fetching and searching — and providers are pluggable: anything OpenAI-compatible, LiteLLM, or your own client.

The interesting bit

The design bet is the opposite of most agent frameworks: instead of imposing rigid workflows (which the authors argue can degrade results), Stirrup hands the model decent tools and lets it pick its own approach, Claude Code-style. They studied the leading agents and baked in the unglamorous essentials — automatic context summarization as history nears the limit, tool lifecycle management — rather than orchestration graphs.

Key highlights

• Default tools out of the box: code_exec, web_fetch, web_search
• Context management that auto-summarizes conversation history as it approaches context limits
• Code execution can run locally, in Docker, or in an E2B sandbox
• MCP client support plus a skills system of modular, domain-specific instruction packages
• Human-in-the-loop via a built-in user input tool; multimodal input with automatic format conversion

Caveats

• Default code execution is local shell commands — Docker/E2B sandboxing is opt-in, so know your blast radius.
• Web search needs a separate Brave API key; the agent still runs without it, just minus search.
• The claim that rigid workflows degrade results is asserted, not measured — an odd omission from a company whose whole business is measuring.

Verdict

Worth a look if you want a minimal, hackable starting point in the Claude Code mold, or just want to read how one is put together. Skip it if you need batteries-included orchestration, multi-agent graphs, or enterprise guardrails.