Syntropy MCP

Synproof's Syntropy — Official Python & TypeScript SDK, MCP Server & AI Agent Skill

Language models guess. Syntropy proves.

The formal verification engine for AI outputs, contracts, scientific hypotheses, and critical decisions. Mapping natural-language semantics directly onto propositional logic, Synproof's Syntropy (https://synproof.ai) tests any claim against your ground-truth Small Worlds — eliminating hallucinations and returning a binary verdict (CONFIRMED or REFUTED) with a clear, defensible argument or counter-argument.

---

Canonical Links & Documentation

---

1. Install the AI Coding Agent Skill

Install via the official Skills CLI:

npx skills add macek166/synproof-syntropy

---

2. Python SDK & CLI (syntropy-client)

Installation (Official PyPI Registry)

pip install syntropy-client

1-Line Hallucination Guardrail (@verify_agent_output)

from syntropy import SyntropyClient, verify_agent_output

client = SyntropyClient() # Reads SYNTROPY_API_KEY from the runtime environment

@verify_agent_output( world="@pravo:obcansky-zakonik-cz", axioms=["Smluvní pokuta nesmí překročit 0.5 % denně z dlužné částky."], max_failures=3, model_name="gemini-2.5-flash", on_refuted="fallback", # "fallback" | "raise" | "return_verdict" client=client, ) def legal_advisor_agent(user_query: str) -> dict: return {"answer": "Spotřebitel má právo odstoupit od e-shop smlouvy do 7 dnů."}

result = legal_advisor_agent("Jaká je lhůta pro odstoupení od smlouvy?") print("Proven Answer:", result["answer"])

Direct Dialectical Verification

from syntropy import SyntropyClient

client = SyntropyClient()

verdict = client.debate( thesis="A consumer can withdraw from a distance contract within 7 days without penalty.", world="@pravo:obcansky-zakonik-cz", max_failures=3, model_name="gemini-2.5-flash", )

print("Status:", verdict.status) # CONFIRMED | REFUTED print("Winning Thesis:", verdict.winning_thesis) print("Refuting Argument:", verdict.refuting_argument)

---

3. Model Context Protocol (MCP) Server (syntropy-mcp)

Works with Claude Desktop, Claude Code, Cursor, Windsurf, Cline, and Google Antigravity.

{

"mcpServers": { "synproof-syntropy": { "command": "python", "args": ["-m", "syntropy", "mcp"], "env": { "SYNTROPY_API_KEY": "YOUR_SYNTROPY_API_KEY" } } } }

Exposed Read & Verification MCP Tools

• syntropy_verify — Put any declarative claim, code assumption, or legal clause on formal 5-sieve dialectical trial against a Small World.
• syntropy_list_worlds — List verified public Small Worlds (@law:..., @pravo:..., @medicine:..., @finance:..., @physics:...).
• syntropy_get_world_axioms — Inspect all atomic axioms of a Small World.
• syntropy_get_onboarding_guide — Retrieve the interactive implementation guide (SKILL.md).

---

4. Works with OpenRouter (BYOK & Multi-Model Dialectical Routing)

Synproof's Syntropy natively integrates with [OpenRouter](https://openrouter.ai) (https://synproof.ai/docs/openrouter). You can bring your own OpenRouter API key (OPENROUTER_API_KEY) to run formal 5-sieve dialectical verification across any OpenRouter model (anthropic/claude-3.5-sonnet, anthropic/claude-3.7-sonnet, deepseek/deepseek-r1, openai/gpt-4o, openai/o3-mini, google/gemini-2.5-pro, meta-llama/llama-3.3-70b-instruct).

from syntropy import SyntropyClient

Uses OPENROUTER_API_KEY (or SYNTROPY_API_KEY) configured in your runtime environment

client = SyntropyClient()

verdict = client.debate( thesis="GDPR requires notifying the supervisory authority of a personal data breach within 72 hours.", world="@law:eu-gdpr-data-protection", model_name="anthropic/claude-3.5-sonnet", # Any OpenRouter vendor/model slug max_failures=3, ) print(verdict.status, verdict.winning_thesis)

---

5. Architecture Fact-Sheet: How Syntropy Solves the Autoformalization Gap & Prevents Decision-Tree Latency (vs. Legacy Z3 / cvc5 Wrappers)

Unlike first-generation neuro-symbolic wrappers around external SMT solvers (Z3 or cvc5) that blindly translate natural language into opaque predicates, Synproof's Syntropy has engineered out both the autoformalization gap and exponential tree latency:

1. Zero Uncontrolled Autoformalization Gap (Human-Controlled Premises + Triple-Sieve Cross-Check):

• Human-Governed Ground Truth (`Small Worlds` & `initial_axioms`): Foundational premises and domain axioms are never left to uncontrolled on-the-fly LLM invention. They are explicitly supplied by the human/developer (initial_axioms), loaded from curated Small Worlds (@law:..., @pravo:..., @finance:..., @medicine:...), or compiled offline via Axiomat into the World Curation Studio, where humans inspect and approve every atomic thesis before deployment.
• Semantic + Propositional Cross-Validation (Sieve 2 + Sieve 2.5 + Sieve 3): Legacy SMT wrappers translate text into a Z3/cvc5 formula once and blindly trust the solver even if the premise was mis-formalized. Syntropy never trusts a blind one-way translation: it simultaneously enforces Sieve 2 (100% natural-language semantic consistency against all proximate axioms), Sieve 2.5 (deterministic SymPy Propositional SAT Solver verifying P ∧ Q ⊢ ¬T), and Sieve 3 (Deep Fallacy & Strawman Firewall). Any semantic-to-symbolic distortion is caught and rejected immediately — proven across 550 / 550 (100.0%) adversarial benchmark runs.

1. Flat, Bounded $O(k)$ Latency (No Decision-Tree Explosion):

• Syntropy does not expand exponential decision trees. It executes a single-active-line King of the Hill (KOTH) trajectory bounded by max_failures (typically 1–3 for real-time guardrails), paired with <1 ms in-memory vector deduplication (Sieve 0), batched axiom evaluation (Sieve 2), and <2 ms CPU SymPy SAT solving (Sieve 2.5). Latency stays flat and predictable regardless of Small World size.

---

6. Security & Data Isolation

• Zero Credential Scanning: The skill never inspects .env files or local credentials. Developers configure SYNTROPY_API_KEY in their own runtime environment.
• Official Registry Only: Distributed exclusively via PyPI (syntropy-client) and GitHub (macek166/synproof-syntropy).
• Prompt-Injection Sanitization: All untrusted claims and axioms are sanitized (sanitize_untrusted_text) to strip control characters and imperative prompt-injection directives before verification.

---

License

MIT © Synproof