io.github.dth0108/bies-guardian

Official3 tools

Deterministic sealed verdicts on public claims and startup ideas (0-LLM claim-safety guardian).

Evaluates public claims and startup ideas for safety without using LLMs.

Captured live from the server via tools/list.

claim_safety_verdict

Deterministic, sealed pre-publication verdict on a PUBLIC claim (marketing copy, product page, ad headline). Returns per-claim verdict (PASS / CONDITIONAL_PASS / STOP), severity, and a reproducible replay_hash. A 0-LLM rules engine grounded in KR 표시광고법 / US FTC claim categories — NOT legal advice, NOT market proof or launch approval. Call this before a public claim ships.

Parameters (3)

copystringrequired
The exact public claim text you intend to publish.
langstring
Language of the claim. Default ko.
marketstring
Regulatory jurisdiction. Default follows lang (ko→KR, en→US).

idea_rehearsal

Deterministic decision-rehearsal verdict on a founder's BUSINESS IDEA (not a public claim). Returns a verdict (Proceed / Conditional GO / Hold·Redesign) with a reproducible seal. Classification is retrieval-grounded; the verdict and seal are reproducible. NOT market proof, NOT investment advice. Use to pressure-test an idea before committing.

Parameters (3)

ideastringrequired
The business idea in the founder's own words — what you sell, to whom.
langstring
Language. Default ko.
marketstring
Market. Default follows lang (ko→KR, en→US).

rejudge_claims

Re-judge a prior claim-safety verdict against new EVIDENCE. Pass the 'artifact' returned by claim_safety_verdict plus 'observations' (what happened). The top-level verdict NEVER changes on evidence (conservative by design); what moves is 4 axes — now-sayable lines (behind your owner-gate), permanently issuer-only claims, claims still short on evidence, and new risks. Deterministic, sealed, stateless (nothing stored).

Parameters (3)

artifactobjectrequired
The full artifact object returned by claim_safety_verdict (the prior verdict being re-judged).
observationsarrayrequired
Evidence events. Each: { outcome: 'paid_yes'|'written_pilot'|'observed_metric'|'rejection', targets: [{ ref_kind: 'claim_id'|'evidence_id', ref: '<id from the artifact>' }], rejection_reason_ref? }.
langstring
Language. Default ko.

Install

claude_desktop_config.json

{
  "mcpServers": {
    "bies-guardian": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://bies.io/mcp"
      ]
    }
  }
}

Desktop config is stdio-only; this bridges via mcp-remote. Native remote: Settings > Connectors.