March 27, 2026 · The Harbor Team
If you've built a Harbor task before, you've probably spent time figuring out the same things everyone else does: multi-container setups, simulated users, adding MCP tools, implementing computer use environments, and more. We built Harbor to make agent evals simple. That’s why we’re releasing the [Harbor Cookbook](https://github.com/harbor-framework/harbor-cookbook). It contains a collection of realistic, ready-to-run examples of how to build evals and optimize agents with Harbor. **What’s inside** We recommend you give your coding agent context on the one closest to what you're building and adapt from there. | Recipe | What it does | | :--------------------------------------------------------------------------------------------------------------------------------- | :--------------------------------------------------------------------------- | | [simple-task](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/recipes/simple-task) | Minimal single-container task | | [multi-container](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/recipes/multi-container) | Docker Compose task where the agent interacts with a locally hosted REST API | | [mcp-tools](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/recipes/mcp-tools) | Giving the agent custom tools via a locally hosted FastMCP server | | [skills](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/recipes/skills) | Recipes for including skills in a Harbor task | | [multi-reward](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/recipes/multi-reward) | Multiple independent verifiers each producing their own score | | [simulated-user](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/recipes/simulated-user) | Agent discovers requirements by talking to a simulated user | | [computer-use-ubuntu](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/recipes/computer-use-ubuntu) | Computer use reference implementation on an Ubuntu virtual desktop | | [computer-use-windows](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/recipes/computer-use-windows) | Computer use reference implementation on a remote Windows desktop (Daytona) | | [dns-blacklisting](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/recipes/dns-blacklisting) | Network-level hostname blacklisting with exact, wildcard, and regex rules | **Beyond evals: optimizing agents** Harbor tasks produce a reward, which means the same datasets you use for evals can also serve as training environments. The cookbook includes two recipes that demonstrate this: one example pairs Harbor with [GEPA](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/gepa) to optimize an agent harness on MedAgentBench. The other is the Harbor integration contributed by [Thinking Machines](https://github.com/harbor-framework/harbor-cookbook/tree/main/harbor_cookbook/tinker_rl), which uses Harbor tasks as RL environments through the Tinker SDK. We welcome feedback on which examples to build next and how to improve Harbor. We’re actively developing the Harbor framework to improve its ability to integrate into optimization loops. **Get started** Clone the cookbook and then run: ```bash tab="uv" theme={"system"} uv tool install harbor ``` ```bash tab="pip" theme={"system"} pip install harbor ``` ```bash theme={"system"} harbor run -p harbor_cookbook/recipes/simple-task -a "