Neura Market
abluva-research/mcp-trust-plane logo

abluva-research/mcp-trust-plane

Free

面向 MCP 的可组合数据安全平面，采集 / 分析 / 防护分层可插拔，覆盖 50+ 企业数据源，为 AI 访问企业数据提供统一的信任与防护层。

FreeFree tier
Type
Open Source
Company
abluva-research
Visit Website GitHub (40 stars)

About abluva-research/mcp-trust-plane

MCP Trust Plane is a pluggable, language-agnostic filter framework for Model Context Protocol (MCP) traffic. It provides a composable data security plane with pluggable layers for collecting, analyzing, and guarding AI access to enterprise data. The framework uses lightweight HTTP services (filters) that intercept MCP tool calls and responses, enforcing security policies such as blocking destructive SQL, redacting PII, throttling abusive agents, capping response sizes, and more. Filters are independent processes that communicate via a well-defined contract: POST /filter (for decision) and GET /health (for readiness). The system supports pre-request and post-response filtering, with actions: allow, block, truncate, or modify. Designed to work with a separate reference gateway and engine, MCP Trust Plane is open source (Apache 2.0) and includes a growing library of common filters covering 50+ enterprise data source providers.

Key Features

Pluggable, language-agnostic filter framework for MCP traffic
Pre-request and post-response filtering with actions: allow, block, truncate, modify
Filters are independent HTTP services with a standard contract (POST /filter, GET /health)
Idempotent and side-effect free filters (except audit/rate limiter state)
Fail-open behavior for slow filters (p95 under 50ms)
Common filters library under Apache 2.0 license
Covers 50+ enterprise data source providers
Open architecture with separate reference gateway and engine

Pros & Cons

Pros
  • Open source and extensible (Apache 2.0)
  • Language-agnostic filters can be written in any language
  • Clear separation of concerns with independent filter processes
  • Well-defined filter contract for easy integration
  • Fail-open design prevents denial of service from slow filters
  • Covers 50+ enterprise providers out of the box
Cons
  • Requires setting up a separate MCP gateway and engine (reference implementations in another repo)
  • Operational complexity of running multiple filter services
  • Still an early-stage project with limited community adoption
  • No built-in support for non-MCP protocols

Best For

Block destructive SQL queries from AI agentsRedact PII before it reaches the modelThrottle abusive or overly frequent agent callsCap response sizes to prevent excessive data exposureRedact sensitive columns from database responsesEnforce enterprise data security policies for AI tool access

FAQ

What is MCP Trust Plane?
MCP Trust Plane is an open-source pluggable filter framework for Model Context Protocol (MCP) traffic. It allows organizations to enforce security guardrails on AI agent access to enterprise data sources by inserting composable filters that can allow, block, truncate, or modify tool calls and responses.
How does MCP Trust Plane work?
An MCP gateway sits in front of the upstream MCP server. For each tool call, the gateway asks an engine to evaluate policies assigned to that target. The engine calls each configured filter via its HTTP endpoint. Filters return an action (allow/block/truncate/modify) and optionally a modified response. The gateway then enforces the decision.
What actions can filters perform?
Filters can respond with allow (let request through unchanged), block (reject with an error message), truncate (return a smaller payload), or modify (replace the payload with an altered version).
What are the performance requirements for filters?
Filter execution time should stay under 50ms p95. The engine treats slow filters as 'allow' to fail-open and avoid blocking legitimate traffic.
Where can I find the reference gateway and engine?
The reference gateway and engine live in a separate repository: secure-ai-plane-x/mcp-filter-gateway. This repository (abluva-research/mcp-trust-plane) ships only the filter implementations and the contract.