rust-pantry

rust-pantry is a store of copy-and-run Rust skeletons and add-on crates, written to be read by an AI agent as much as by a person. Four templates cover a command-line tool, an axum server with a shared wire-type crate, a Tauri desktop app with a React front end, and a native egui app. Nine recipes add one capability each, such as a command registry that makes an app agent-ready over MCP, an embedded Claude Code or Codex assistant behind an approval gate, background jobs, settings with migrations, signed self-update, or a service wrapped behind a trait with a fake. The catalog in AGENTS.md gives every item a pick-when and a skip-when line, so an agent can match the app in front of it and take only what it needs.

• –Four templates (cli, server, tauri, egui): each builds, lints and tests on its own; server, tauri and egui keep their logic in a UI-free core or protocol crate
• –new.sh generator: copies a template, renames the myapp placeholders in file contents and paths, runs git init, and prints each chosen recipe's wiring steps
• –Agent-ready recipe: one command registry callable from the UI, a CLI, a token-gated loopback JSON-lines channel and an MCP server
• –Embedded-agent recipe: runs the user's own Claude Code or Codex CLI as the app's assistant, with the app's commands as its only tools and every call through an approval gate
• –Desktop and service recipes: background jobs with cancel, versioned settings with migrations, single-instance forwarding, opaque file handles, argon2 and session auth, and Ed25519-signed self-update
• –Providers recipe: wraps an external service behind a trait with a deterministic fake and a demo mode, so tests and CI run without network or accounts
• –Shared conventions in every template: unsafe code forbidden, clippy -D warnings in CI, thiserror in libraries and anyhow in binaries and one pinned toolchain, plus, in the multi-crate templates, a cargo xtask layers check that keeps UI dependencies out of core crates
• –Root CI that checks every template, the recipes workspace on Linux and Windows, and a project generated by new.sh from each template with all the recipes that fit it

• –Writing a catalog an agent can act on: every item needs honest pick-when and skip-when lines, or the agent takes recipes speculatively and leaves the app with code nobody asked for
• –Keeping recipes adoptable into an existing app, where the host's error types, logging and async runtime win over the recipe's own choices
• –Running an embedded agent safely: the assistant gets only the app's own commands, and a missing answer, a dropped reply or a cancelled turn must all mean deny
• –Making a self-updating binary trustworthy: a validly signed manifest for the wrong app or platform must still be rejected

Each recipe is a small crate with a recipe.toml that carries its pick-when and skip-when lines, and a README that explains how it works, how to wire it in and how to adopt it into an existing app. The ai recipe starts Claude Code with its built-in tools off and only the app's own MCP server, refuses to start Codex unless a check before each start shows its dangerous features and outside MCP servers off, and routes every tool call through one gate where an unanswered approval always means deny. Self-update verifies the Ed25519 signature over the raw manifest bytes before anything is parsed, then checks app, platform and version before it downloads and swaps the executable. A cargo xtask layers check and a root CI that generates a project from every template keep the store itself from drifting.

A private repository that is planned to be made public. A new Rust app can start from a template that builds, lints and tests on its own and take only the recipes it needs, with CI generating and checking a project from every template to keep them working.