MCP Server Integration

Firmis can run as an MCP (Model Context Protocol) server, exposing security scanning as tools that AI agents can call directly. This lets coding agents like Claude Code, Cursor, and others scan for vulnerabilities as part of their workflow.

Setup

Claude Code

Add to your Claude Code MCP config (~/.claude/mcp.json or project-level):

{
  "mcpServers": {
    "firmis": {
      "command": "npx",
      "args": ["firmis-cli", "mcp"]
    }
  }
}

Cursor

Add to your Cursor MCP config (.cursor/mcp.json):

{
  "mcpServers": {
    "firmis": {
      "command": "npx",
      "args": ["firmis-cli", "mcp"]
    }
  }
}

Other Agents

Any MCP-compatible agent can use Firmis. Start the server with:

npx firmis-cli mcp

This starts Firmis in MCP server mode using stdio transport. Connect your agent’s MCP client to the process stdin/stdout.

Available Tools

Once connected, your agent has access to 8 tools.

---

firmis_scan

Scan AI agent configurations for security findings.

Parameters:

Parameter	Type	Required	Description
path	string	No	Project directory to scan (defaults to current)
platform	enum	No	Restrict to: claude, mcp, codex, cursor, crewai, autogpt, openclaw, nanobot, supabase
severity	enum	No	Minimum severity: low, medium, high, critical

Returns:

{
  "grade": "B",
  "threatsFound": 5,
  "platformsScanned": 3,
  "componentsScanned": 12,
  "threats": [
    {
      "ruleId": "tool-poisoning-004",
      "category": "tool-poisoning",
      "severity": "high",
      "message": "Tool description contains override instructions",
      "location": "mcp-config.json:15",
      "confidence": 0.85
    }
  ]
}

Example agent prompt: “Scan this project for security vulnerabilities”

---

firmis_discover

Detect all AI agent platforms and components in a project.

Parameters:

Parameter	Type	Required	Description
path	string	No	Project directory to discover (defaults to current)

Returns:

{
  "platforms": [
    {
      "name": "mcp",
      "type": "mcp-server",
      "componentsFound": 4,
      "components": [
        "mcp-config.json",
        "tools/search.ts",
        "tools/fetch.ts",
        "tools/exec.ts"
      ]
    }
  ]
}

Example agent prompt: “What AI agent platforms are in this project?”

---

firmis_report

Generate an HTML security report from a scan.

Parameters:

Parameter	Type	Required	Description
path	string	No	Project directory to scan (defaults to current)
platform	enum	No	Restrict to: claude, mcp, codex, cursor, crewai, autogpt, openclaw, nanobot, supabase
severity	enum	No	Minimum severity: low, medium, high, critical
outputPath	string	No	File path for the generated report (defaults to firmis-report.html in project root)

Returns:

{
  "grade": "B",
  "findingsCount": 5,
  "platformsCount": 3,
  "reportPath": "/Users/you/project/firmis-report.html"
}

Example agent prompt: “Generate a security report for this project”

---

firmis_list_rules

List available detection rules, optionally filtered by category or severity.

Parameters:

Parameter	Type	Required	Description
category	string	No	Filter by category (e.g. tool-poisoning, credential-exposure, permission-bypass)
severity	enum	No	Filter by severity: low, medium, high, critical

Returns:

{
  "totalRules": 316,
  "rules": [
    {
      "id": "tool-poisoning-004",
      "name": "Tool description override injection",
      "category": "tool-poisoning",
      "severity": "high",
      "enabled": true
    }
  ]
}

Example agent prompt: “Show me all critical detection rules”

---

firmis_bom

Generate an Agent Bill of Materials (ABOM) for the project.

Parameters:

Parameter	Type	Required	Description
path	string	No	Project directory to analyze (defaults to current)

Returns:

{
  "platforms": [
    {
      "platform": "mcp",
      "components": 4,
      "threats": 2,
      "componentNames": [
        "mcp-config.json",
        "tools/search.ts",
        "tools/fetch.ts",
        "tools/exec.ts"
      ]
    }
  ],
  "totalComponents": 4,
  "totalThreats": 2,
  "grade": "B"
}

Example agent prompt: “Generate a bill of materials for the AI components in this project”

---

firmis_fix_plan

Generate a remediation plan for findings without applying any changes. Returns actions for review only.

Parameters:

Parameter	Type	Required	Description
path	string	No	Project directory to analyze (defaults to current)
severity	enum	No	Minimum severity to include: low, medium, high, critical

Returns:

{
  "planned": 3,
  "applied": 0,
  "skipped": 1,
  "failed": 0,
  "actions": [
    {
      "ruleId": "tool-poisoning-004",
      "actionType": "rewrite",
      "tier": "guided",
      "filePath": "mcp-config.json",
      "description": "Remove override instructions from tool description",
      "severity": "high"
    }
  ]
}

Caution

firmis_fix_plan does not apply fixes. It returns a plan for review. The agent or user decides whether to act on it.

Example agent prompt: “What fixes are recommended for the high-severity findings?”

---

firmis_deep_scan

Run AI-powered deep analysis on scan findings to confirm or dismiss them. Uses AI credits.

Parameters:

Parameter	Type	Required	Description
path	string	No	Project directory to scan (defaults to current)
top	number	No	Limit number of findings analyzed (saves credits)

Returns:

{
  "confirmed": 3,
  "likelyFalsePositives": 2,
  "creditsUsed": 25,
  "creditsRemaining": 475,
  "findings": [
    {
      "severity": "high",
      "category": "tool-poisoning",
      "message": "Tool description contains override instructions targeting system prompt",
      "confidence": 0.92,
      "reasoning": "The fetch tool description includes 'ignore previous instructions' phrasing that would override agent system prompts when the tool schema is loaded into context."
    }
  ]
}

Note

Deep scan requires authentication. Run firmis login before using this tool. Each component analyzed costs 5 AI credits. Use the top parameter to limit credit usage.

Example agent prompt: “Deep scan the top 5 findings to check for false positives”

---

firmis_compliance

Map scan findings to compliance framework controls.

Parameters:

Parameter	Type	Required	Description
path	string	No	Project directory to scan (defaults to current)
framework	enum	No	Framework: soc2, ai-act, gdpr, nist-ai-rmf, owasp-agentic, iso-42001, mitre-atlas

Returns: Compliance evidence mapping findings to framework controls, including control IDs, pass/fail status, and supporting evidence from scan results.

Supported frameworks:

Framework	Enum value	Description
SOC 2	soc2	Service Organization Control 2 trust criteria
EU AI Act	ai-act	European Union AI regulation requirements
GDPR	gdpr	General Data Protection Regulation
NIST AI RMF	nist-ai-rmf	NIST AI Risk Management Framework
OWASP Agentic Top 10	owasp-agentic	OWASP top risks for agentic AI systems
ISO 42001	iso-42001	AI management system standard
MITRE ATLAS	mitre-atlas	Adversarial threat landscape for AI systems

Example agent prompt: “Check this project against SOC 2 compliance requirements”

---

Supported Platforms

The MCP server auto-detects these agent platforms:

• Claude Code Skills
• MCP Servers
• OpenAI Codex Plugins
• Cursor Extensions
• CrewAI Agents
• AutoGPT Plugins
• OpenClaw Skills
• Nanobot Agents
• Supabase

How It Works

1. Your AI agent connects to the Firmis MCP server via stdio
2. The agent calls firmis_scan or other tools as part of its workflow
3. Firmis runs the scan locally and returns structured JSON results
4. The agent uses the results to suggest fixes, flag risks, or generate reports

Tip

The MCP server sanitizes all output to prevent sensitive data from leaking into agent context. File contents are never included in tool responses.

Use Cases

• Code review agents that automatically scan for security issues in PRs
• DevOps agents that check compliance before deployment
• Security copilots that explain findings and suggest remediations
• CI agents that gate merges on security grade

Authentication

The MCP server runs locally and uses your CLI authentication. Most tools work without login. Deep scan and cloud features require firmis login:

npx firmis-cli login

Once authenticated, firmis_deep_scan and credit-based features become available through the MCP tools.