Toulmini enforces rigorous argumentation through Stephen Toulmin's model. It's an MCP server that forces LLMs into structured reasoning—no external API keys required.

The problem: LLMs hedge, skip logical steps, and avoid self-critique.

The solution: Toulmini enforces falsifiable claims, sequential phases, and mandatory adversarial testing. Weak logic triggers circuit breakers that terminate the analysis.

Query → Data/Claim → Warrant/Backing → Rebuttal/Qualifier → Verdict
         Phase 1        Phase 2            Phase 3          Phase 4

pip install toulmini

claude mcp add toulmini -- python -m toulmini.server

{
  "mcpServers": {
    "toulmini": {
      "command": "python",
      "args": ["-m", "toulmini.server"]
    }
  }
}

toulmini-cli --verify          # Health check
toulmini-cli --config          # Show current config
toulmini-cli --install -       # Print MCP snippet

"Analyze this argument: Should we allow human genetic engineering?"

Toulmini executes all phases automatically and returns a verdict.

If the warrant or backing is weak, the chain terminates immediately—no weak arguments reach the verdict phase.

Convene expert perspectives before Phase 2 or 3:

consult_field_experts(
    query="Should we allow genetic engineering?",
    perspectives=["Bioethicist", "Medical Geneticist", "Disability Rights Advocate"]
)

Citation reliability: Without web search, citations come from LLM training data and may be outdated or hallucinated. Treat them as leads to investigate, not verified sources.

Failure modes:

• Weak warrant/backing → chain terminates at Phase 2
• Absolute rebuttal → verdict must be "overruled"
• Confidence < 30% → verdict should be "overruled" or "remanded"

git clone https://github.com/Hmbown/Toulmini.git
cd Toulmini
pip install -e ".[dev]"
pytest

See CONTRIBUTING.md for guidelines.

• Documentation
• Examples
• Hegelion — Dialectical reasoning (sister project)