crosscheck-mcp

A multi-LLM MCP server. Ask several models the same question, debate them, orchestrate them, audit them, or rubric-grade their work — all from a single MCP tool surface that drops into Claude Desktop / Claude Code / Cursor / Continue / any other MCP host.

This is the TypeScript implementation. There is also a Python implementation in ../python/ — both ship the same tool surface; pick whichever fits your stack. The TypeScript build runs natively in Node 18.17+ and has no Python dependency at runtime.

What it does

Twenty-six tools across deliberation, planning, auditing, and operations. Highlights:

• confer / debate / coordinate / triangulate — parallel panel calls, structured deliberation, structured synthesis with dissent
• audit — rubric-graded scoring; single-judge or multi-judge consensus (median + std-dev disagreement detection, severity-aware obvious-failure flags)
• orchestrate — planner-driven DAG of LLM subtasks with optional cheap-mode tier-aware routing
• create / create_cheap — lifecycle macros: scope → build → review → audit, with retry-on-failure
• solve — iterative propose → verify → retry with sandboxed shell verifiers
• verify — deterministic property checks (text, shell, url_head)
• recall / session_memory / scoreboard / explain — operational introspection

Cross-cutting features: scoreboard-driven router for panel selection, session circuit breakers (cost / tokens / wall / DAG breadth), cross-provider canary detection for indirect prompt-injection, early-stop on agreement, prompt canonicalisation cache, structured claims with supports/attacks edges.

Install

npm install -g crosscheck-mcp

This installs the crosscheck-mcp binary on your PATH.

The package needs at least one LLM provider API key in the environment. Supported providers (set any subset):

export ANTHROPIC_API_KEY=...
export OPENAI_API_KEY=...
export XAI_API_KEY=...
export GEMINI_API_KEY=...
export MISTRAL_API_KEY=...
export GROQ_API_KEY=...
export DEEPSEEK_API_KEY=...

Use it with an MCP host

Claude Code

Add the server to your project's .claude/settings.json (or the user-level config):

{
  "mcpServers": {
    "crosscheck": {
      "command": "crosscheck-mcp",
      "env": {
        "ANTHROPIC_API_KEY": "${ANTHROPIC_API_KEY}",
        "OPENAI_API_KEY":    "${OPENAI_API_KEY}"
      }
    }
  }
}

Restart Claude Code; the mcp__crosscheck__* tools become available.

Tip — type xc or XC at the start of a prompt to route it to crosscheck; the server ships an MCP instructions payload that teaches Claude Code the convention. Details in ../../CROSSCHECK_USAGE.md.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "crosscheck": {
      "command": "crosscheck-mcp",
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "OPENAI_API_KEY":    "sk-..."
      }
    }
  }
}

Restart Claude Desktop. Tools appear under the hammer icon.

Cursor

Cursor reads MCP servers from ~/.cursor/mcp.json. Same shape:

{
  "mcpServers": {
    "crosscheck": {
      "command": "crosscheck-mcp",
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "OPENAI_API_KEY":    "sk-..."
      }
    }
  }
}

Any other MCP host

Spawn crosscheck-mcp over stdio. It speaks MCP JSON-RPC 2.0 — no custom protocol.

Optional configuration

Environment variables:

The Python backstop is OPTIONAL. Every v1 feature path on every tool runs natively in TypeScript. The bridge is useful for:

1. Cross-language parity testing (the test/parity/ fixtures check that the TS implementation produces byte-equal output against the Python oracle).
2. The reactive opt on orchestrate (mid-flight DAG mutation), which still defers to bridge.
3. A few remaining advanced opts that no longer have a Node-side implementation but kept the bridge as a fallback when host integrations aren't wired (e.g. worker_tools without innerCallers).

Everything in the Tools list above works without `CROSSCHECK_BRIDGE_PYTHON` being set.

Observability

Every tools/call request emits one structured tool_invoke event through a pluggable event bus. The default sink is ND-JSON to stderr — MCP hosts already capture stderr into their session logs, so this lights up out of the box with no configuration.

Event shape

{
  "event":          "tool_invoke",
  "id":             "1a3c0f5d",
  "tool":           "confer",
  "started_at":     "2026-06-06T19:42:11.231Z",
  "duration_ms":    1842,
  "status":         "ok",
  "args_keys":      ["question", "providers", "untrusted_input"],
  "result_keys":    ["answers", "question", "tool"],
  "envelope_bytes": 12451
}

Failure paths emit the same shape with status: "error" and an error_message (truncated to 512 chars). When a tool returns a domain-level error envelope (with error_code), the call still completes — status stays "ok" and error_code is mirrored to the event for easy filtering:

{
  "event": "tool_invoke", "id": "...", "tool": "audit",
  "status": "ok",
  "error_code": "AUDIT_NO_AUDITOR",
  "...": "..."
}

What's deliberately not in the event: argument values, response text, transcripts. Long prompts + multi-LLM debates would dwarf the signal and may carry sensitive content. Key inventory + envelope size is usually enough to triage a regression; pull the full envelope from a tool replay when you need it.

Routing

CROSSCHECK_EVENTS_PATH is required when CROSSCHECK_EVENTS is file or both. Parent directories are created on first emit.

Custom emitters (programmatic)

import { setEventEmitter } from "crosscheck-mcp";

setEventEmitter({
  emit(event) {
    metrics.histogram("crosscheck.tool_duration_ms",
      event.duration_ms, { tool: event.tool, status: event.status });
  },
});

The emitter is replaced for the whole process; for test isolation use the RecordingEmitter export and restore the original via getEventEmitter() / setEventEmitter() in beforeEach / afterEach.

Programmatic use

import { createServer, connectAndServe } from "crosscheck-mcp";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";

const server = createServer({ /* providers, storage, pricing, repoRoot ... */ });
await server.connect(new StdioServerTransport());

Development

git clone https://github.com/fxspeiser/crosscheck-agent
cd crosscheck-agent/servers/typescript
npm install
npm run build        # tsup → dist/{node-stdio,node-http,browser-ext}.{js,cjs}
npm test             # vitest — 1,175 tests
npm run typecheck    # tsc --noEmit
npm run docs:usage   # regenerate the routing-convention doc from src/instructions.ts

End-to-end stdio handshake:

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"x","version":"0"}}}' \
  '{"jsonrpc":"2.0","method":"notifications/initialized"}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"ping","arguments":{"echo":"hello"}}}' \
  | node dist/node-stdio.js

License

PolyForm Noncommercial 1.0.0. Commercial licenses are tracked in ../../COMMERCIAL-LICENSES.md; contact frank@metafide.io for inquiries.