Skip to content

MCP_SERVER.md

Latest commit

 

History

History
166 lines (119 loc) · 5.62 KB

MCP_SERVER.md

File metadata and controls

166 lines (119 loc) · 5.62 KB

MCP Server — Connect agent-bom to AI Assistants

agent-bom exposes 36 security tools as an MCP server. Any MCP-compatible client can connect and get vulnerability scanning, blast radius analysis, compliance checks, and supply chain verification through natural conversation.

See also:

Quick Start

Claude Desktop

Add to your claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/):

{
  "mcpServers": {
    "agent-bom": {
      "command": "agent-bom",
      "args": ["mcp", "server"]
    }
  }
}

Restart Claude Desktop. You can now ask: "Scan my AI agents for vulnerabilities"

Claude Code

If you already use the Claude CLI, add agent-bom directly:

claude mcp add agent-bom -- uvx agent-bom mcp server

Claude Code project-level MCP servers are also discovered from ~/.claude.json.

Cortex CoCo

Add to ~/.snowflake/cortex/mcp.json:

{
  "mcpServers": {
    "agent-bom": {
      "command": "uvx",
      "args": ["agent-bom", "mcp", "server"]
    }
  }
}

CoCo can then call the same 36 agent-bom tools over MCP.

agent-bom also discovers Cortex auxiliary security files alongside mcp.json:

  • settings.json
  • permissions.json
  • hooks.json

Cursor / Windsurf / VS Code

Add to your MCP settings (.cursor/mcp.json or equivalent):

{
  "mcpServers": {
    "agent-bom": {
      "command": "agent-bom",
      "args": ["mcp", "server"]
    }
  }
}

SSE Transport (Remote / Multi-Client)

agent-bom mcp server --sse --host 0.0.0.0 --port 8000

Connect any SSE-capable MCP client to https://your-server/sse. For remote deployments, put SSE behind TLS and authentication at your proxy or ingress.

Docker

docker run -it --rm \
  -v ~/.config:/root/.config:ro \
  agentbom/agent-bom:latest mcp server

Runtime proxy

Use the proxy when you want to inspect or enforce on MCP traffic between a client and a third-party server:

agent-bom proxy "npx @modelcontextprotocol/server-filesystem /workspace"

This keeps the real server behind agent-bom and enables runtime detectors for tool drift, credential leakage, injection patterns, sequence risk, and related policy decisions.

For JSON-configured clients like Claude Desktop or Cortex CoCo, use:

agent-bom proxy-configure --log-dir ~/.agent-bom/logs --detect-credentials

Add --apply to write the wrapped config back to compatible JSON MCP config files.

Tool Categories (36 tools)

Category Tools What They Do
Scan scan, code_scan, vector_db_scan, gpu_infra_scan, ai_inventory_scan Discover agents, scan packages, code, vector stores, GPU infra, and AI usage
Check check, verify, marketplace_check, license_compliance_scan Pre-install CVE gate, integrity verification, marketplace trust, and license policy
Blast Radius blast_radius Map CVE → package → MCP server → agent → credentials → tools
Registry registry_lookup, inventory, where, fleet_scan Query the MCP registry, inspect discovery paths, and summarize fleet inventories
Compliance compliance, cis_benchmark, aisvs_benchmark Run OWASP, NIST, MITRE ATLAS, CIS, and AISVS-aligned posture checks
Policy policy_check, remediate Evaluate policies and generate guided remediation plans
Inventory inventory List agents/servers without CVE scanning
Trust marketplace_check, runtime_correlate, tool_risk_assessment Score package trust, correlate runtime usage, and assess live tool capability risk
Skills skill_scan, skill_verify, skill_trust Instruction-file trust, provenance, and tool-poisoning detection
Graph / Runtime context_graph, graph_export, runtime_correlate, tool_risk_assessment Visualize lateral movement, export graph data, and connect runtime logs to findings
AI supply chain dataset_card_scan, training_pipeline_scan, browser_extension_scan, model_provenance_scan, prompt_scan, model_file_scan, ingest_external_scan Scan AI artifacts, prompts, model files, browser extensions, and external scanner results

Example Conversations

"Are my AI agents vulnerable?"

Agent-bom discovers your Claude Desktop, Cursor, and VS Code MCP configs, extracts all server packages, queries OSV/NVD for CVEs, and shows the blast radius chain.

"Is it safe to install mcp-server-sqlite?"

Runs pre-install check: CVE scan, typosquat detection, OpenSSF Scorecard, license analysis, and supply chain provenance verification.

"Show me my compliance posture"

Runs OWASP LLM Top 10, MITRE ATLAS, NIST AI RMF, and CIS benchmarks against your infrastructure. Returns per-framework pass/fail/warn.

Security Model

  • Read-only: Only List/Describe/Get operations. Zero write calls.
  • No credential storage: Never stores, logs, or transmits your credentials.
  • No network exfiltration: Scans local configs, queries public CVE databases.
  • Agentless: No agents installed on targets.

Resources

The server exposes two MCP resources:

  • registry://servers — Browse the full 427+ server security metadata registry
  • policy://template — Default security policy template

Prompts

Built-in prompts for common workflows:

  • security-scan — Full agent + MCP server vulnerability scan
  • pre-install-check — Check a package before installing
  • compliance-posture — Multi-framework compliance assessment