Guides and reference for authoring and running recipes.
Recipes are an open package format for Pi agents. An agent recipe is a Git directory that configures a complete agent: instructions, model, tools, skills, connections, and the agents it orchestrates. Pi runs it. A host supplies credentials, storage, and lifecycle.
Quick start
There are two entry points, and they produce the same directory.
Start with Introspection
Install the CLI, then let setup prepare Pi, Recipes, and the Introspection plugin for the coding agents it detects. Setup shows its plan and asks before changing anything.
npm install --global @introspection-ai/cli
introspection setup
Open the repository where the recipe should live in Codex or Claude Code and describe the agent you need. The plugin creates or migrates the recipe with you, proves it in fresh Pi sessions, and returns the command for running it locally. A new task discovers the installed workflow automatically.
Build me an agent that <describe your agent>
Start with Pi
Clone a recipe that already runs, read it, then change the parts where your work is different.
pi install npm:@introspection-ai/recipes # once per machine
git clone https://github.com/introspection-recipes/template-claude
pi --recipe ./template-claude
How Recipes work
- Recipes the package: what is in the directory, and how it is read.
- Agents identity, model, tools, inheritance, and delegation.
- Skills and prompts your method, and the shape of what comes out.
- MCP two policy gates, CLI and tools mode, deferred tools.
- Extensions tools you write in TypeScript, owned by the package.
- Interactions ask the user something, in a terminal or over your protocol.
- Evals and judges Evalite, Harbor, and judges you calibrate locally.
Deploy it
- Deploying what changes when a host supplies credentials and endpoints.
- Telemetry pass a tracer, get GenAI spans for every model and tool call.
- SDK reference resolve a recipe and construct its Pi session from your code.