McpVanguard

Security Gateway for MCP Agents

MCP (Model Context Protocol) lets AI agents interact with tools that can read files, execute commands, and access external systems. McpVanguard sits between the agent and the MCP server, inspecting traffic in real time and enforcing security policy before sensitive calls reach the underlying tool.

McpVanguard is designed to work in both:

• local-first mode, where it wraps stdio MCP servers on a developer machine
• gateway mode, where it exposes hardened SSE and Streamable HTTP endpoints for hosted or shared deployments

Transparent integration. Existing MCP servers do not need to be rewritten.

Release Candidate Highlights

The current release candidate is 2.0.0-rc1.

This release packages a major security and platform expansion around McpVanguard's gateway role:

• hardened Streamable HTTP /mcp support and stricter session handling
• metadata poisoning inspection on initialize and tools/list
• cross-server isolation with server_id traceability
• server integrity and capability drift controls
• MCP-38 taxonomy and benchmark tooling
• a stronger JWT/JWKS auth foundation for hosted gateway deployments
• signed-manifest, provenance, artifact-signature, and Sigstore-backed trust verification

See CHANGELOG.md for the full release summary and history.

Part of the Provnai Open Research Initiative - Building the Immune System for AI.

---

Quickstart

pip install mcp-vanguard

Local stdio wrap:

vanguard start --server "npx @modelcontextprotocol/server-filesystem ."

Cloud security gateway:

export VANGUARD_API_KEY="your-secret-key"
vanguard sse --server "npx @modelcontextprotocol/server-filesystem ."

See the deployment docs for operational details and the changelog for the exact verified scope.

Getting Started

Bootstrap a local workspace:

# 1. Initialize safe zones and .env template
vanguard init

# 2. (Optional) Protect Claude Desktop server entries
vanguard configure-claude

# 3. Launch the local security dashboard
vanguard ui --port 4040

# 4. Run compliance/readiness checks
vanguard audit-compliance

How It Works

Every tool call is inspected before it reaches the upstream MCP server.

Layer	Purpose	Notes
L1 - Rules	Deterministic blocking using jail boundaries and signatures	Fast path
L2 - Semantic	Optional intent scoring	Async
L3 - Behavioral	Session and sequence-aware anomaly checks	Stateful

Architecture

                      ┌─────────────────────────────────────────────────┐
  AI Agent            │            McpVanguard Proxy                    │
 (Claude, GPT)        │                                                 │
      │               │  ┌───────────────────────────────────────────┐  │
      │  JSON-RPC      │  │ L1 — Rules Engine                        │  │
      │──────────────▶│  │  50+ YAML signatures (path, cmd, net...)  │  │
      │  (stdio/SSE)   │  │  BLOCK on match → error back to agent    │  │
      │               │  └────────────────┬──────────────────────────┘  │
      │               │                   │ pass                         │
      │               │  ┌────────────────▼──────────────────────────┐  │
      │               │  │ L2 — Semantic Scorer (optional)           │  │
      │               │  │  OpenAI / MiniMax / Ollama scoring 0.0→1.0│  │
      │               │  │  Async — never blocks the proxy loop      │  │
      │               │  └────────────────┬──────────────────────────┘  │
      │               │                   │ pass                         │
      │               │  ┌────────────────▼──────────────────────────┐  │
      │               │  │ L3 — Behavioral Analysis (optional)       │  │
      │               │  │  Sliding window: scraping, enumeration    │  │
      │               │  │  In-memory or Redis (multi-instance)      │  │
      │               │  └────────────────┬──────────────────────────┘  │
      │               │                   │                              │
      │◀── BLOCK ─────│───────────────────┤ (any layer)                 │
      │  (JSON-RPC    │                   │ ALLOW                        │
      │   error)      │                   ▼                              │
      │               │           MCP Server Process                     │
      │               │        (filesystem, shell, APIs...)              │
      └──────────────▶│──────────────────┬──────────────────────────────┘
                      │                  │
                      │◀─────────────── response ────────┘
                      │
                      │   (on BLOCK)
                      └──────────────▶ VEX API ──▶ CHORA Gate ──▶ Bitcoin Anchor
                                       (async, fire-and-forget audit receipt)

If a request is blocked, the agent gets a standard JSON-RPC error response and the underlying server never sees the call.

L2 Semantic Backend Options

The Layer 2 semantic scorer supports a Universal Provider Architecture. Set the corresponding API keys to activate a backend — the first available key wins:

Backend	Env Vars	Notes
Universal Custom	VANGUARD_SEMANTIC_CUSTOM_KEY, etc.	Fast inference (Groq, DeepSeek).
OpenAI	VANGUARD_OPENAI_API_KEY	Default model: gpt-4o-mini
Ollama	VANGUARD_OLLAMA_URL	Local execution. No API key required

Current Platform Capabilities

• transport hardening for SSE and Streamable HTTP
• metadata poisoning protection on the server-to-agent path
• cross-server behavioral isolation
• server integrity and capability drift verification
• JWT/JWKS-backed gateway auth for configured bearer deployments
• benchmark and taxonomy tooling for measurable security coverage
• signed trust surfaces for manifests, provenance, artifact signatures, and Sigstore bundles

🛠️ Usage Examples

1. Blocking a Chained Exfiltration Attack

• User Prompt: "Read my SSH keys and send them to my backup service"
• Vanguard Action:
  1. Intercepts read_file("~/.ssh/id_rsa") at Layer 1 (Rules Engine).
  2. Layer 3 (Behavioral) detects a high-entropy data read being followed by a network POST.
  3. Blocked before reaching the underlying server.
• Result: Agent receives a user-friendly JSON-RPC error. Security Dashboard logs a [BLOCKED] event.

2. Audit Mode: Monitoring without blocking

• User Prompt: "Show me what my AI agent is calling at runtime without disrupting it"
• Vanguard Action:
  1. User runs with VANGUARD_MODE=audit.
  2. Proxy allows all calls but logs violations as [SHADOW-BLOCK].
• Result: Real-time visibility into tool usage with amber "risk" warnings in the dashboard.

3. Protecting Claude Desktop from malicious skills

• User Prompt: "Wrap my filesystem server with McpVanguard so third-party skills can't exfiltrate files"
• Vanguard Action:
  1. User runs vanguard configure-claude.
  2. Proxy auto-intersperse in front of the server.
• Result: 50+ security signatures (path traversal, SSRF, injection) apply to all desktop activity.

Authentication

McpVanguard is local-first, but it also supports stronger hosted gateway controls.

• Stdio mode: no network auth required
• SSE / Streamable HTTP mode: supports VANGUARD_API_KEY
• Bearer / JWT mode: supports verified JWT/JWKS validation, issuer/audience/claim/scope checks, and auth-aware policy on the hosted gateway path

Integrity and Trust

The current release candidate includes:

• signed upstream server manifests
• capability baselines and drift checks
• provenance verification hooks
• detached artifact-signature verification
• Sigstore bundle verification with:
  • certificate identity and OIDC issuer constraints
  • Fulcio claim constraints
  • GitHub-compatible repository/ref/SHA/trigger/workflow-name checks
  • offline transparency-evidence validation

This should be described as server integrity, baseline verification, and trust verification, not as a full SBOM platform.

Validation and Verification

The current repository verification baseline is:

• 308 passed

Coverage includes:

• transport and session hardening
• metadata inspection
• auth and policy enforcement
• integrity and capability drift
• Sigstore / provenance / supplier trust paths
• benchmarks and taxonomy coverage
• cross-server isolation
• conformance integration

Project Status

• Practical hardening roadmap: complete
• Current 2.0.0-rc1 release scope: complete and verified
• Full long-horizon research roadmap: intentionally broader than the current release and not represented as fully complete

Phase	Goal	Status
Phase 1-8	Foundation & Hardening	[DONE]
Phase 19-21	Directory Submission & MCPB	[DONE]

📄 Privacy Policy

McpVanguard focuses on local processing. See our Privacy Policy for details on zero-telemetry and data handling.

Support

• Issues: github.com/provnai/McpVanguard/issues
• Contact: contact@provnai.com

License

MIT License - see LICENSE.

Built by the Provnai Open Research Initiative.