MCP 服务器

@palveron/mcp-server

Provides advisory AI-governance checks for MCP hosts and coding agents, enabling policy verdicts, PII masking, and audit traces before executing tool calls.

README

<p align="center"> <h1 align="center">@palveron/mcp-server</h1> <p align="center">Official Palveron MCP server — advisory AI-governance checks for MCP hosts and coding agents</p> </p>

⚡ Early access. Palveron is launching soon and access is currently invite-gated. This server is fully functional — you just need a pv_live_ key, which comes with an Early Access invite. Join the waitlist → palveron.com/early-access

Give your MCP-capable client — Claude Code, Cursor, any other MCP host — governance tools it can call before executing an action: policy verdicts, PII-masking results, and an audit trace for every check, all from the Palveron AI Governance Gateway.

Drop-in MCP server — one binary, configure your MCP host, you're done
Advisory governance verdicts — the gateway decides allow / mask / deny / hold; the agent (or your wrapping code) acts on the verdict
Audit trail per check — trace IDs flow back to the Palveron dashboard
Built on @palveron/sdk — retry, circuit breaker, typed errors out of the box

Advisory check vs. enforced governance — which one do you need?

This package is the advisory path. It cannot physically stop a tool call: it answers the question "is this allowed?", and the calling agent (or your code via check()/wrap()) is expected to honor the answer. That is by design — it works with any MCP host, needs no network topology change, and gives you verdicts plus audit traces immediately.

Enforced, non-bypassable governance for MCP tool calls is a different product surface: the Palveron remote MCP proxy. There you register the target MCP server in the Palveron dashboard and point your IDE at POST {gateway}/api/v1/mcp/proxy/{server_id} instead of the target — every tools/call then structurally passes through policy checks, approval holds, and drift detection before it can reach the target server. No cooperation from the agent required.

	This package (`@palveron/mcp-server`)	Remote MCP proxy
Model	Voluntary pre-flight check	In-path enforcement
Can a call bypass it?	Yes — enforcement depends on the caller honoring the verdict	No — the proxy sits between client and target server
Setup	`npx`, one env var	Register the target server in the dashboard, point the IDE at the proxy URL
Use it when	You want verdicts + audit traces inside any MCP host, or wrap tool calls programmatically	You need governance that agents cannot skip

See the MCP setup guide for the enforced path.

Installation

npm install -g @palveron/mcp-server
# or use directly via npx
npx @palveron/mcp-server

Quick Start

The pv_live_… key in the examples below requires an Early Access invite — request one here. Once you have a key, everything works out of the box.

Claude Code

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "palveron": {
      "command": "npx",
      "args": ["@palveron/mcp-server"],
      "env": {
        "PALVERON_API_KEY": "pv_live_xxx",
        "PALVERON_BASE_URL": "https://gateway.palveron.com"
      }
    }
  }
}

Other MCP hosts

Any MCP-capable host can launch the server via the palveron-mcp binary:

PALVERON_API_KEY=pv_live_xxx palveron-mcp

Tools exposed via MCP

Tool	Purpose
`palveron_check`	Check a tool call against governance policies before execution. Returns the decision (`PASSED`, `BLOCKED`, `MODIFIED`/`REDACTED`/`ANONYMIZED` with the sanitized input, `PENDING_APPROVAL`, …), whether the call may proceed, findings, and the audit trace ID.
`palveron_status`	Governance status and diagnostics — gateway connectivity, circuit-breaker state, configured local lists.
`palveron_policy_list`	List the project's active policies so the host can explain its decisions.

Configuration

The server reads its configuration from environment variables. The MCP host sets them; you never put secrets in code.

Variable	Required	Description
`PALVERON_API_KEY`	yes	API key (`pv_live_…`) — comes with an Early Access invite
`PALVERON_BASE_URL`	no	Override the gateway endpoint (default: `https://gateway.palveron.com`)
`PALVERON_FAIL_OPEN`	no	`true` = allow tool calls when the gateway is unreachable (default: `false`, fail-closed)
`PALVERON_LOG_LEVEL`	no	`debug` / `info` / `warn` / `error` / `silent` (default: `info`)
`PALVERON_ALWAYS_BLOCK`	no	Comma-separated tool names that are always denied locally, without a gateway call
`PALVERON_ALWAYS_ALLOW`	no	Comma-separated tool names that always pass locally, without a gateway call