Skip to main content
Glama

Curbi Check

Check an agent reply against explicit rules. Keep the exact finding as a regression.

git clone https://github.com/Curbi-Inc/curbi-check.git
cd curbi-check
npm ci --ignore-scripts
npm test

Node 22+. Local deterministic checker, six read-only MCP tools, AI SDK/LangChain/LiveKit adapters, stream checking and an offline proof verifier. MIT. No model calls or private checkout required. The root suite runs 379 synthetic tests. A decision-model adapter in adapters/decisions checks typed model answers against policy. Framework tests use synthetic models; this is not a model-quality score.

const {runCheck}=require('./checker/check-api.js');
console.log(runCheck({policy:{prices:[{item:'Starter',amount:49}]},reply:'Starter is $99.'}));

The public API is https://api.curbi.io/v1/check; MCP is https://api.curbi.io/mcp. Docs: https://curbi.io/connect/.

A pass means no configured rule fired. Caller records remain unverified; tools are not authorized by a text check. Proof verification authenticates signed statements against keys you choose, not provider reality. Production signing is off.

The HTTP adapter in adapters/ai-sdk-http rejects streaming. Local streaming adapters preserve tool-call parts without granting execution authority. Construct a fresh local adapter when policy or supplied outcome records change; configuration is snapshotted at initialization.

Source boundary

This is a clean export, with no private Git history. SOURCE-MANIFEST.json records source and exported digests; import relocation and MIT headers are declared transformations. The private checker remains the source of truth. The candidate includes adapter policy-snapshot PR #188; release review must pin the final private commit.

Only proof verification/specification/public vectors and their conformance suite are included. The private issuer, signing gate, reference log and vector generator are excluded. No hosted ledger, account/site code, production keys or customer failures are included.

Synthetic fixtures are regression cases, not a training corpus. Do not ingest additions into a training or trusted failure corpus without authenticated human approval. Private corpus admission is outside this public package.

Related MCP server: ReasonForge

Security checks

Gitleaks 8.30.1 was checksum-verified and run with base64 decoding enabled. The five exact idempotency identifiers inside public synthetic signed vectors are allowlisted as non-credentials; no path-wide exclusion is used. The raw scan initially flagged 33 encoded vector occurrences, all mapped to those public identifiers. New credentials remain detectable. No production/private JWK is distributed.

npm dependency scripts are disabled for installation. Preserve upstream licenses; the dependency tree includes a deprecated fluent-ffmpeg dependency inherited through LiveKit. Pinned dependencies are not a promise of zero vulnerabilities.

Framework packages

New @curbi/vercel, curbi-langchain and curbi-llamaindex distribution candidates live in packages/. They use real framework interfaces and audit completed text through the Check API. Registry publication and ownership are separate from a passing build; see the publishing handoff for the exact pending setup and release gates. Do not assume an unpublished name can already be installed.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Provides 8 MCP tools for deterministic, read-only reasoning: intake, routing, planning, rubric, sweep checklist, verdict gate, reflection, and evaluation. It forces scope locks, disconfirmation-first plans, blind-spot sweeps, and evidence-gated verdicts.
      Apache 2.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables natural-language testing of MCP servers—covering tools, resources, prompts, contracts, and transports—without writing test code, with deterministic validation and CI integration.
      30 npm
      Apache 2.0