AutoContext — Self-Improving Agent Harness

Main

AutoContext 更像 harness 层，而不是一份单独 prompt。它的核心工作是围绕场景做循环评估，把改进结果继续反馈给下一轮 agent。

这对已经过了玩具阶段的团队很关键：你需要的是可重复改进，而不是靠经验猜测某次运行为什么成功。

README 给出了两条可执行集成路径：基于 Pi 或 provider 环境变量的 CLI 路径，以及 Claude Code 里的 MCP 路径。这说明它面向的是实际工作流，而不只是 demo。

Source-backed notes

• README offers a 30-second path with uv tool install autocontext==0.5.0.
• It documents provider-based operation for Anthropic, OpenAI, Gemini, Mistral, Groq, OpenRouter, Azure, Claude CLI, Codex CLI, and MLX.
• Pi runtime and Claude Code MCP integration are both called out as supported paths.

FAQ

问：Is AutoContext tied to one model provider? 答：No. The README documents several providers plus Claude CLI, Codex CLI, and MLX paths.

问：What is the fastest install path? 答：uv tool install autocontext==0.5.0, then point it at Pi or another supported provider.

问：Why use it? 答：It helps repeated agent runs improve systematically instead of starting from scratch every time.