hardHats

coldHat is for play; hardHats are for work. hardHats is the framework — the factory that makes crews. A crew is a line-up of agents, each wearing a hat, run across a federation of free model backends.

A hat is one of two things:

• an LLM hat — a model wearing an instruction (a role, a method, an output rule), or
• a gate hat — a deterministic function: a real check, a linter, a measurement.

The same engine runs both, so a crew can put an LLM drafter and a deterministic verifier on the same line. That is the point: the LLM hats do the judgement, the gate hats do the facts, and nobody has to trust a model for something a rule can decide.

A crew is config, not code

from hardhats import Hat, Crew

crew = Crew("grant", [
    Hat("researcher", model=("groq", "openai/gpt-oss-20b"),
        instruction="List what a Statement of Need must establish."),
    Hat("writer", model=("github", "openai/gpt-4o"),
        instruction="Draft it in a cold, plain register. No AI-tell words."),
    Hat("critic", gate=slopgate),          # a deterministic gate hat
])

# pre-flight the crew against a conflict graph before it runs
conflicts = crew.preflight(load("dataset/prompt-conflicts.json"))

outputs = crew.run("Write a Statement of Need for an accessibility tool.")

Each hat sees the task and every prior hat's output, down the line.

The federation

hardhats.backends speaks to one chat() over many OpenAI-compatible (and Gemini) backends — Gemini, Groq, OpenRouter, Ollama, plus GitHub Models, Cloudflare Workers AI, Cerebras, Mistral, NVIDIA NIM, SambaNova and Cohere — each read from the environment, keys never printed. Add a key, gain a perspective.

Crews in the family

• a11yhat — the WCAG accessibility machine, four crew leads as deterministic gate hats.
• a grant crew, a compliance crew, a prompt-check crew — each a few lines of config.
• the bench — 22 keyword-summonable crews (an expert for every call, each with a real gate). See hardhats/crews/README.md; list them with python -m hardhats.crews.

Tests

python -m unittest discover -s tests — the engine (pipeline + pre-flight) is verified with gate hats only, so it proves out with zero model calls.

License

MIT. Copyright Vincent Gonzalez / F-Keys Creative LLC.