Stirrup

Stirrup's central design bet is that the model, not the framework, should decide how to finish a task. The docs state this directly: Stirrup "gets out of the way and lets the model choose its own approach," similar to Claude Code, and argues that many frameworks impose rigid workflows that can degrade results. That shows up in the mechanics. Create an Agent without passing tools and it loads default_tools() — web search, web fetching, and local code execution. Custom tools are a four-part class (name, description, Pydantic parameter model, executor), with sync executors run in a separate thread by default and async executors supported natively. ToolProvider handles tools needing connections or temp directories, and metadata like ToolUseCountMetadata is aggregated across a run so you can track how often each tool fired and attach your own cost counters via the Addable protocol. Context handling is handled for you: the framework summarizes conversation history as it approaches context limits, which matters for long-running agent loops. Multimodal input covers images, video, and audio with automatic format conversion, and tools can return images back to the model through ImageContentBlock. Provider flexibility is real rather than nominal — the Quick Start uses OpenRouter via ChatCompletionsClient with anthropic/claude-opus-5, the docs show DeepSeek through its OpenAI-compatible endpoint, and LiteLLMClient covers Anthropic and Google with a model_slug. The honest weaknesses are the ones the project itself implies. Stirrup is early-stage and the documentation, while structured, is thinner than what mature frameworks ship. Nearly everything needs external API keys (OpenRouter, Brave Search, plus whatever coding-backend credentials you choose) and Python fluency. There is no GUI, no enterprise support or SLA, and nothing resembling the integration breadth of LangChain. Optional components like Docker, E2B, MCP, and browser tooling are extras you install separately, so your first hour involves deciding which ones you actually need. Where it fits: engineers who want to understand what their agent is doing, teams that have outgrown the rigidity of a heavy orchestration library, and researchers comparing provider behavior across the same tool loop. Where it doesn't: non-programmers, teams that need a supported product with a contract, and any project whose main requirement is a large catalog of pre-built connectors.