Cooper: Okay — this is a framework for packaging AI agent capabilities like installable skills. Imagine pip, but for machine skills, not libraries.
Miles: Uh… sure. But we already have MCP and Skyet already shipped agent skills as server specs. What’s actually new?
Cooper: Right, but—every project reinvents tool definitions, system prompts, and safety rules. Skillware declares a skill as a bundle: executable Python, cognition instructions, governance rules, and a standardized tool schema. One install, one manifest, works across Gemini, Claude, Ollama, GPT, Llama.
Miles: So it’s a package manager for agent capabilities.
Cooper: Exactly. apt-get for skills — apt-get apt-get for skills — you literally pip install it.
Miles: Okay, okay. Show me the thing.
Cooper: …fine. Let’s spin it up. First, install: pip install skillware. If you’re on Gemini, grab gemini extra too: pip install "skillware" — space between the brackets matters, or it errors, and I just spent twenty minutes debugging that.
Miles: Naturally. Of course it’s a one-character typo that swallows the whole afternoon.
Cooper: Once installed, skillware list prints your locally available skills and confirms path resolution.
Miles: And you trust whatever random skill you installed?
Cooper: Well… that’s the fine print. The loader pulls the bundle — manifest.yaml, skill.py, instructions.md, card.json — and adapts it to the model’s tool format. But yes, every skill runs in your process and can read every variable in os.environ. There’s a whole trust model section in the docs for a reason.
Miles: Hm. So the trust surface is everything the skill’s maintainer says it is.
Cooper: Pretty much. You’re still on the hook for authorship and keys — skill keys get declared in the Skill library per bundle. The framework itself doesn’t ship keys.
Miles: I see. So it’s an App Store with no review board?
Cooper: I mean — it’s open source, so the review board is open pull requests and crickets at 3 AM on a Tuesday. But yeah, the roadmap page is titled ‘Vision’ and it does picture an App Store for agents.
Miles: Stop—so you’re fundraising on GitHub stars?
Cooper: Eighty-seven downloads this week and counting! I’ll take it. Eighty-seven happy little agents doing real work.
Miles: …I did not expect that to land as hard as it did. Fine. Walk me through a real loop.
Cooper: Model Context Protocol is server specs; Skillware is code-first — the skill is executable Python, not a Dockerfile. The loader adapts the manifest and instructions to whatever provider tool format you’re using.
Miles: So if I swap from gemini to Ollama, I just change the provider setup and the rest stays?
Cooper: That’s the pitch. The tool call comes back, you run skill.execute with the args, then feed the result back into the model. The gemini usage guide has the exact three-step loop.
Miles: Okay. Okay, this is actually clean.
Cooper: I know right? Clean enough that I want to build a skill for cursor’s router output. Fewer tiny review comments about the same log field — that’s a product win.
Miles: …So this is basically our control-infrastructure thesis with a hoodie on.
Cooper: Exactly. The hoodie makes it fun.
Miles: I’m still uneasy about the trust model.
Cooper: Read the trust model docs. Per-skill key isolation and the whole git-clone-and-audit thing. It’s not perfect, but it’s a hell of a lot better than copy-pasting the same prompt template into every repo.
Miles: Fair. Fine by me as long as the docs don’t live in a gated PDF only.
Cooper: No gate. SECURITY.md is right there in the repo. You can grep for ‘env’ and see every place os.environ shows up.
Miles: …I don’t know, sometimes I think you’d trust a skill just because it has the right name.
Cooper: Oh come on! I just spent twenty minutes debugging a pip install typo — I am NOT installing random finance skills willy-nilly!
Miles: Fine. I’ll concede the demo works. What’s the path to getting one new skill into the registry?
Cooper: Open a PR, follow the template, include deterministic tests, and you’re in. The contributing guide walks you through it — scoped PRs, deterministic behavior, and verified tests.
Miles: Mmm. So the App Store part is real, but the review is purely mechanical?
Cooper: For now. Vision page talks about graduated curation, but today it’s PRs and CI.