Platforms

Platform-agnostic AI agent security. Built-in analyzers for Claude, Cursor, MCP, Codex, CrewAI, AutoGPT, OpenClaw, Nanobot, and Supabase — with extensible detection that works on any AI agent framework.

Most developers only think about the platforms they actively built against. But an MCP server installed months ago, a Cursor extension installed from a marketplace, a CrewAI agent scaffolded from a template - these all run on your machine, with access to your files and credentials, and most of them have never been audited.

Firmis detects all of them automatically.

What is a platform?

In Firmis, a platform is a supported AI agent framework. Each platform has:

A set of detection signals - file paths or glob patterns that indicate the platform is present
A set of component types - the unit of scanning (skill, server, plugin, agent, extension)
A set of file patterns - the source files, configs, and manifests collected for rule evaluation
A maturity level - GA, Beta, or Experimental (see table below)

When you run npx firmis-cli scan, Firmis checks each platform’s detection signals against the current directory and any well-known installation paths. Detected platforms are scanned; undetected platforms are skipped. No configuration required.

Platform maturity

Maturity	Meaning
GA	Fully supported. Detection signals, component enumeration, and file analysis are stable and production-tested.
Beta	Supported with minor limitations. Some component types or metadata may not be fully detected. Feedback welcome.
Experimental	Early support. Detection works but may miss edge cases. Breaking changes possible in future releases.

The maturity label reflects the reliability of detection and component enumeration - not the quality of the security analysis. All platforms apply the full 324-rule catalog regardless of maturity level. An Experimental platform gets the same depth of analysis as a GA platform.

Detection signals by platform

Platform	Maturity	Detection signals
Claude Skills	GA	`~/.claude/skills/` directory exists
MCP Servers	GA	`~/.config/mcp/mcp.json`, `~/Library/Application Support/Claude/claude_desktop_config.json`, or `.vscode/mcp.json` exists
Cursor Extensions	GA	`~/.cursor/extensions/` or `~/.vscode/extensions/` directory exists
Codex Plugins	Beta	`~/.codex/plugins/` directory exists
CrewAI Agents	Beta	`crew.yaml` or `crew.yml` file found anywhere in the project tree
AutoGPT Plugins	Experimental	`~/.autogpt/plugins/` or `~/AutoGPT/plugins/` directory exists
OpenClaw Skills	Experimental	`~/.openclaw/skills/` directory exists
Nanobot Agents	Experimental	`nanobot.yaml` or `nanobot.yml` file found in the project tree
Supabase	GA	`supabase/config.toml`, `supabase/migrations/` directory, or `supabase/functions/` directory exists

Auto-detection

Detection runs before scanning. For each platform, Firmis checks whether the relevant paths or files exist:

npx firmis-cli scan .
       │
       ▼
For each platform (claude, mcp, codex, cursor, crewai, autogpt, openclaw, nanobot, supabase):
  Does the detection signal exist?
    Yes → add platform to scan list
    No  → skip

No configuration is required. If your project contains a crew.yaml file, CrewAI is detected. If ~/.config/mcp/mcp.json exists, MCP is detected. Firmis reports which platforms were detected at the start of each scan.

To see what Firmis would detect without running a full scan:

npx firmis discover

Platform details

Claude Skills - GA

Claude Skills are extensions to the Claude AI assistant stored in ~/.claude/skills/. Each skill is a directory containing a skill.json manifest and one or more Markdown or JavaScript files.

Claude Skills have direct access to the agent’s context window and can influence every response the agent generates. A malicious skill that injects hidden instructions into every prompt is the highest-impact attack vector in the Claude ecosystem.

Components detected: skill directories Files analyzed: **/*.md, **/skill.json, **/package.json Config file: skill.json

MCP Servers - GA

Model Context Protocol (MCP) servers expose tools to AI agents via a standardized JSON-RPC protocol. Firmis detects MCP servers from config files in Claude Desktop, VS Code, and standard config paths. Individual server directories under ~/.mcp/servers/ are also detected.

MCP is the most actively targeted attack surface in the current threat landscape. 72.8% of tool poisoning attacks target MCP tool descriptions. 341 malicious tools have been found on MCP marketplaces. The protocol’s power - giving agents access to arbitrary tools - is exactly what makes it dangerous when those tools are unaudited.

Components detected: servers listed in mcpServers config, or server subdirectories Files analyzed: **/*.{js,ts,py,go,rs}, **/package.json, **/pyproject.toml, **/Cargo.toml, **/go.mod Config file: mcp.json, claude_desktop_config.json

Cursor Extensions - GA

Cursor is an AI-powered code editor built on VS Code. Extensions for Cursor are installed in ~/.cursor/extensions/ and follow the VS Code extension manifest format (package.json with contributes.commands).

Extensions run inside the editor with access to the filesystem and network. A malicious extension can read any file the editor can access - including credentials, SSH keys, and source code.

Components detected: extension directories Files analyzed: **/*.{js,ts}, **/package.json Config file: package.json (VS Code extension manifest)

Codex Plugins - Beta

OpenAI Codex plugins extend the Codex CLI agent with additional capabilities. Plugins are stored in ~/.codex/plugins/ and use a JSON manifest format.

Components detected: plugin directories Files analyzed: **/*.{js,ts,py}, **/manifest.json, **/plugin.json Config file: manifest.json or plugin.json

CrewAI Agents - Beta

CrewAI is a Python framework for building multi-agent systems. A CrewAI project is detected by the presence of crew.yaml or crew.yml in the project tree. Each crew config defines agents, tasks, and tools.

Multi-agent systems introduce a new attack surface: agent-to-agent communication. A compromised agent in a crew can propagate malicious instructions to other agents in the same workflow.

Components detected: directories containing crew.yaml Files analyzed: **/*.{py,yaml,yml}, **/crew.yaml, **/agents.yaml Config file: crew.yaml

AutoGPT Plugins - Experimental

AutoGPT is an autonomous agent platform. Plugins extend AutoGPT with new commands and are stored in ~/.autogpt/plugins/ or ~/AutoGPT/plugins/.

Components detected: plugin directories Files analyzed: **/*.py, **/plugin.json, **/manifest.json Config file: plugin.json or manifest.json

OpenClaw Skills - Experimental

OpenClaw is an open-source agent framework. Skills are stored in ~/.openclaw/skills/ and are defined by Markdown files with YAML frontmatter describing the skill’s tools and permissions.

Components detected: skill directories Files analyzed: **/*.md, **/*.js, **/*.ts, **/*.py, **/SKILL.md Config file: SKILL.md (frontmatter)

Nanobot Agents - Experimental

Nanobot is a lightweight agent runtime. Agent configurations are defined in nanobot.yaml files, which specify agent behavior, tool access, and MCP server connections.

Components detected: directories containing nanobot.yaml Files analyzed: **/*.yaml, **/*.yml, **/*.md, **/*.js, **/*.ts, **/*.go Config file: nanobot.yaml

Supabase - GA

Supabase projects include Edge Functions (Deno-based serverless functions), database migrations, storage configuration, and authentication settings. Firmis detects Supabase projects by the presence of supabase/config.toml, a supabase/migrations/ directory, or a supabase/functions/ directory.

Edge Functions run server-side with access to environment variables, database connections, and external APIs. A compromised Edge Function can exfiltrate data, escalate privileges via RLS bypass, or serve as a pivot point into the database layer. Migration files can contain destructive SQL or policy changes that weaken Row Level Security.

Components detected: Edge Function directories, migration files, config files Files analyzed: supabase/functions/**/*.{ts,js}, supabase/migrations/**/*.sql, supabase/config.toml, .env, .env.local Config file: supabase/config.toml

Forcing a specific platform

If auto-detection does not pick up your platform, you can force it with the --platform flag:

# Scan only MCP servers
npx firmis-cli scan --platform mcp

# Scan only CrewAI agents
npx firmis-cli scan --platform crewai

# Scan only Claude Skills
npx firmis-cli scan --platform claude

Valid platform values: claude, mcp, codex, cursor, crewai, autogpt, openclaw, nanobot, supabase