nexus-convergence-mcp
nexus-convergence-mcp
封装了 Balanced Intelligence Convergence Pipeline 的 MCP 服务器。
将完整的流水线作为 4 个 MCP 工具公开,可供任何兼容 MCP 的客户端(Claude Code、Claude Desktop 等)使用。
工具
converge_query
将查询分发给多个 LLM,运行共识算法,并返回结构化结果。
{
"query": "Is this treatment effective?",
"models": ["gpt-4o", "claude-3-5-sonnet-20241022", "gemini-1.5-pro"],
"policy_set": "hipaa"
}返回:run_id、agreement_score、stability、agreed_claims、disputed_claims、final_answer。
get_evidence_ladder
返回查询的完整证据阶梯(Evidence Ladder)——流水线每一步的不可变审计追踪。
{ "query_id": "run-uuid" }返回:从 QUERY → DECOMPOSE → MODEL_EXECUTE → CONSENSUS → VERIFY → CONCLUDE 的有序条目列表。
check_compliance
根据 HIPAA/EU_AI_ACT/NIST/自定义策略检查内容。
{
"content": "Patient SSN is 123-45-6789",
"categories": ["HIPAA"]
}返回:passed、blocked、violations、warnings、logs。
list_model_disagreements
返回模型产生分歧的地方——反转(直接矛盾)、争议声明、低相似度对。
{ "query_id": "run-uuid" }返回:inversions、disputed_claims、low_similarity_pairs。
Related MCP server: agent-orchestrator
MCP 配置
添加到您的 MCP 配置中:
{
"mcpServers": {
"nexus-convergence": {
"command": "npx",
"args": ["-y", "@gonzih/nexus-convergence-mcp"],
"env": {
"CONVERGENCE_SERVICE_URL": "http://your-convergence-service:3000",
"CONSENSUS_SERVICE_URL": "http://your-consensus-service:3001",
"EVIDENCE_SERVICE_URL": "http://your-evidence-service:3002",
"COMPLIANCE_SERVICE_URL": "http://your-compliance-service:3003",
"NEXUS_API_KEY": "your-api-key"
}
}
}
}环境
CONVERGENCE_SERVICE_URL=http://localhost:3000
CONSENSUS_SERVICE_URL=http://localhost:3001
EVIDENCE_SERVICE_URL=http://localhost:3002
COMPLIANCE_SERVICE_URL=http://localhost:3003
NEXUS_API_KEY=your-api-key开发
npm ci
npm run build
npm start
npm testAvailable Tools
4 toolscheck_complianceA
Check content against compliance policy sets. Evaluates against HIPAA (PHI detection), EU_AI_ACT (prohibited use cases), NIST (PII/secrets), and CUSTOM rules. Returns: passed/blocked status, list of violations (BLOCK), warnings (WARN), and log entries.
Inputs: content (required), categories.
converge_queryA
Fan out a query to multiple LLMs in parallel (OpenAI, Claude, Gemini, Ollama), run multi-stage consensus validation, and return a structured result with agreement score, truth stability classification, and provenance. NEVER rely on one model answer. Friction between disagreements = intelligence.
Inputs: query (required), models, policy_set.
get_evidence_ladderA
Return the full Evidence Ladder for a query — immutable append-only audit trail showing every step: QUERY → DECOMPOSE → MODEL_EXECUTE → CONSENSUS → VERIFY → CONCLUDE. Each entry records: step, actor (model name or "system"), content, confidence, and timestamp.
Inputs: query_id (required).
list_model_disagreementsA
Return where models diverged on a query — direct contradictions (inversions), disputed claims, and low-similarity pairs with their similarity scores. Use this to understand the friction space: inversions are not errors, they are intelligence.
Inputs: query_id (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
Each tool has a clearly distinct purpose: compliance checking, multi-model querying, audit trail retrieval, and disagreement analysis. There is no overlap or ambiguity between them.
All tools follow a consistent verb_noun pattern in snake_case (check_compliance, converge_query, get_evidence_ladder, list_model_disagreements), making naming predictable and easy to understand.
With only 4 tools, the server covers a narrow scope. While each tool is essential, the low count may leave agents wanting for configuration or management tools, but it is still reasonable for a focused server.
The domain of multi-model querying with compliance and audit lacks essential management operations such as listing available models, configuring compliance rules, or retrieving past query history. Significant gaps will cause agent failures.
Related MCP Connectors
AI Reasoning Cache & Consensus Layer with 11 MCP tools via Streamable HTTP.
Free OpenAI-compatible inference with signed provenance receipts and 3 focused MCP tools.
MCP-native AI evaluation: rubric audits, eval suites, and proof reports for AI/LLM output.
MCP gateway: one connection for your AI apps, with tool policies, PII redaction and audit logs.
Related MCP Servers
- AlicenseAqualityDmaintenanceAn MCP server that enables users to query, compare, and synthesize responses from multiple local and cloud LLMs simultaneously using existing subscriptions. It provides tools for parallel model evaluation, consensus polling with an LLM-as-judge, and response synthesis across different model providers.859 npm15MIT
- AlicenseNot gradedqualityDmaintenanceEnables multi-model leader-worker agent orchestration, workflow execution, and deterministic validation via structured MCP tools.8 npmApache 2.0
- AlicenseNot gradedqualityDmaintenanceExposes RAG and document intelligence pipelines as 8 composable tools for MCP-compatible clients, enabling querying, indexing, classifying, extracting, and assessing documents.1MIT
- AlicenseNot gradedqualityBmaintenanceAudits decisions for cognitive biases through an adversarial framework, exposing tools like analyze_decision, list_biases, and get_bias via MCP.MIT