Summary
OakShelf is the main active bet: a package system for distributing reusable capabilities across AI coding agents.
It gives skills, tools, workflows, and profiles a versioned path from source files to an installed agent environment. Packages can be validated, bundled, published, resolved, verified, and linked into supported agents such as Claude Code and Codex.
The point is not another prompt library. OakShelf treats agent behavior as a package with declared entrypoints, dependencies, versions, integrity checks, and a lockfile.
Direction
The next version is trying to make agent capabilities portable without hiding how they work.
The core lifecycle should remain inspectable:
package source
-> validate and bundle
-> publish and resolve
-> install and verify
-> link into an agentInstallation should not execute package code. Integrity checks should prove that installed files match the published bundle. Human approval still decides whether a package is trustworthy enough to use.
The project should work across agent environments while preserving each agent's native file conventions.
Current proof
The current proof of concept includes a registry, resolver, command-line interface, deterministic package bundles, semantic version resolution, digest verification, dependency handling, and adapters for Claude Code and Codex.
Example packages exercise complete workflows rather than isolated prompts. They show how conversational skills can compose with deterministic tools and policy files without giving the agent ownership of every decision.
The source and technical walkthrough live in the OakShelf repository.
Open questions
- What package metadata helps a person evaluate trust before installation?
- Which entrypoint conventions can work across different agent environments?
- What belongs in a conversational skill, and what must remain deterministic code?
- How should organizations publish private packages without creating a second system?
- Which examples make the package model clear before the ecosystem is large?