Skill Container – a decent way to run and distribute agent skills

Name: Skill Container – a decent way to run and distribute agent skills
Availability: InStock
Author: observerw

by observerw·Feb 15, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainNiche Gem

OCI-based agent skill packaging, but limited adoption and niche audience versus established agent frameworks.

Strengths

•Identifies real gaps in existing agent skill formats (Agent Skills): portability, safety boundaries, reproducibility, and distribution are genuine engineering problems.
•OCI container approach elegantly solves portability, dependency isolation, and supply-chain security in one move—genuine engineering insight.
•Clear spec with worked examples and rationale; thoughtful GitHub-native distribution workflow.

Weaknesses

•Competing against established agent ecosystems (Claude's native format, Cursor's integrations, Anthropic's own tooling). Adoption depends on ecosystem buy-in, not just technical merit.
•Very early: 1 star, no releases, no existing skill library to demonstrate viability or attract users.

Post Description

Hi HN — I’m trying to figure out a simple/open way to package agent skills with a clear runtime contract and a sane distribution/update story. What I found frustrating in many existing skill formats is that they’re basically lack of portability: - unclear how to run (Python shebangs, hidden deps, host-specific assumptions) - easy to pollute local environments and drift across machines/versions - little to no safety boundary And distribution is often hand-wavy too: no standard install/update/uninstall flow, no integrity/pinning story.

I found that simply introducing OCI containers solves pretty much all of this in one move.

So, Skill Container’s approach is: OCI containers + GitHub distribution. - Each skill is a repo: SKILL.md + Containerfile + a real CLI entrypoint (e.g. cli.py) + deps in pyproject.toml - Authors publish images to GHCR (ghcr.io/...); users “install” by cloning the repo - Running is just docker run --rm ... ghcr.io/<owner>/<skill>:<tag> --help with explicit mounts; --help is the discovery surface - Updates are predictable: git pull for docs + docker pull for runtime (keeping docs and behavior in sync)

Any opinions would be appreciated!