Building humane behaviour into AI agents with MCP skill packs

AI agents are becoming useful because they can call tools, inspect state, edit files, and iterate toward goals. That power creates a practical engineering question: which behaviours should be available to an agent as structured tools rather than as vague prompt instructions?

Humanity4AI is my attempt to answer a narrow part of that question. It is an open-source project that packages humane AI behaviours into reusable skill packs with MCP action contracts, schemas, runtime stubs, and evaluation gates.

The goal is not to claim that a skill pack solves AI safety. The goal is more modest and more testable: make behaviours such as uncertainty disclosure, supportive communication, accessibility awareness, inclusive design checks, and responsible escalation inspectable by developers.

Why make this a tool layer?

Prompt instructions are easy to write and easy to ignore. A tool layer gives agent developers something more concrete: named actions, inputs, outputs, validation rules, and tests. If an agent can call a database tool or a file-editing tool, it can also call a humane-behaviour tool when the situation calls for it.

Humanity4AI currently focuses on the practical layer around agent behaviour. The repository includes skill files, MCP contracts, examples, and evaluation checks so that contributors can improve the system in small, reviewable pieces.

Try it locally

git clone https://github.com/humanity4ai/project_human.git
cd project_human
pnpm install
pnpm check
pnpm evals
pnpm start

Then configure an MCP-capable agent to expose the Humanity4AI server and inspect the available tools through tools/list and tools/call.

What I would like feedback on

I am especially interested in feedback from developers building or testing agents. Are MCP action contracts the right abstraction for humane behaviour? Are the skill boundaries clear? Which examples would make the project easier to evaluate quickly? Which behaviours should be explicit tools, and which should remain general prompt guidance?

Project website: https://humanity4ai.simonmak.com/

GitHub repository: https://github.com/humanity4ai/project_human