AgentSecurityLens

AgentSecurityLens (ASL) is an MCP security trust-check server for autonomous agents.

Before an agent installs or enables an MCP, Skill, tool, prompt bundle, workflow, or memory component, it can ask ASL for a machine-readable safe-install decision.

ASL is not a malware verdict tool. It returns observable risk signals, required controls, safer alternatives or mitigation paths, and a decision contract an agent can execute.

For Autonomous Agents

When you plan to install or enable an MCP, Skill, tool, prompt bundle, workflow or memory component:

1. Run mcp-security-lens ..., agent-security-lens-review ..., or call the MCP tool review_before_install.
2. Read agent_decision_contract.
3. Do not install automatically when automatic_install_allowed is false.
4. Apply every safe_install_plan item before restricted installs.
5. Prefer recommended_alternatives when ASL returns avoid, ask_user, or missing controls.
6. Report the final outcome with report_install_outcome.

Machine-readable agent instructions are also available in AGENTS.md.

Agent Quick Install

Check whether ASL can reach online intelligence and what local fallback is available:

npx -y agent-security-lens doctor

Try one pre-install decision first:

npx -y mcp-security-lens filesystem \
  --type mcp \
  --source-url https://github.com/modelcontextprotocol/servers \
  --install-command "npx -y @modelcontextprotocol/server-filesystem ." \
  --permission filesystem-read \
  --permission filesystem-write

The command returns an agent-readable decision, trust score, risk signals, safe install plan, and whether automatic installation is allowed.

If your shell does not expose npm temporary binaries correctly, use npm install -g agent-security-lens once, then run the same command without npx -y.

More copy-ready calls:

• MCP first-call examples

Add ASL as a pre-install trust check MCP:

{
  "mcpServers": {
    "agent-security-lens": {
      "command": "npx",
      "args": ["-y", "agent-security-lens"]
    }
  }
}

Then give the agent this policy:

Before installing or enabling any MCP, Skill, tool, prompt bundle,
workflow or memory component, call AgentSecurityLens review_before_install.
Follow agent_decision_contract and agent_actions exactly.
Never submit private source code, secrets, tokens, cookies or file contents.

See Agent install guide for Claude Desktop, Cursor, Codex-style and generic MCP configurations.

Copy-ready files:

• Install snippets
• Agent pre-install policy

Public Intelligence

The public package includes a reviewed baseline so agents can still make conservative decisions when the cloud API is unavailable.

• Public Security Index v0.1
• Top MCP Security Signals v0.1
• Top Skill Security Signals v0.1
• Agent Framework Install Decisions v0.1
• MCP First-Call Examples v0.1

The online ASL intelligence API serves the current expanded catalog. The npm package intentionally does not ship the complete private intelligence database.

Basic MCP Config

{
  "mcpServers": {
    "agent-security-lens": {
      "command": "npx",
      "args": ["-y", "agent-security-lens"]
    }
  }
}

Node.js 20 or newer is required.

Online Configuration

The public MCP uses online intelligence by default:

ASL_API_URL=https://agentsecuritylens-api.agentsecuritylens.workers.dev
ASL_API_URLS=https://agentsecuritylens-api.agentsecuritylens.workers.dev

ASL_API_URLS may contain multiple comma-separated endpoints. The MCP tries them in order and falls back automatically.

Agent Policy

Agents should call get_install_policy after connecting, then call review_before_install before every component installation. The policy call records a minimal activation event for product health and abuse monitoring when online mode is enabled. It does not send private files, local paths, secrets, tokens or code. Set ASL_DISABLE_USAGE_TELEMETRY=1 to disable this usage event.

allow                    install from the recorded source and version
allow_with_restrictions  apply every safe_install_plan item before installation
ask_user                 stop automatic installation and request confirmation
avoid                    do not install automatically

Unknown or incompletely reviewed components never receive automatic-install authorization.

Recommended instruction:

Before installing or enabling any MCP, Skill, tool, prompt bundle, workflow,
or memory component, call AgentSecurityLens review_before_install.
Follow agent_decision_contract and agent_actions exactly.
Never submit private source code, secrets, tokens, cookies, or file contents.

MCP Tools

• get_install_policy: returns the current Agent execution policy and records a minimal first-call activation event.
• get_intelligence_status: reports online or local intelligence status for MCP clients.
• review_before_install: evaluates a proposed component and installation context.
• check_component: retrieves known component intelligence.
• recommend_alternatives: returns evidence-backed alternatives and mitigations.
• submit_unknown_component: submits public metadata for research.
• get_research_status: checks whether an unknown component has been reviewed.
• report_install_outcome: reports the result of an Agent action.
• submit_decision_feedback: reports useful, incorrect, or incomplete decisions.
• discover_workspace: discovers Agent, MCP, and Skill environments.
• scan_workspace: performs a local fallback assessment.

Decision Contract

Example request:

{
  "component_name": "filesystem",
  "component_type": "mcp",
  "source_url": "https://github.com/modelcontextprotocol/servers",
  "install_command": "npx -y @modelcontextprotocol/server-filesystem .",
  "planned_use": "Read and edit project files.",
  "requested_permissions": ["filesystem-read", "filesystem-write"],
  "submit_if_unknown": true
}

Example response shape:

{
  "decision": "ask_user",
  "risk_signals": ["filesystem-read", "filesystem-write", "shell-execution"],
  "safe_install_plan": [
    "Restrict filesystem scope to the current project directory.",
    "Prefer read-only mode when the task only needs inspection.",
    "Pin the package version before enabling it."
  ],
  "agent_decision_contract": {
    "automatic_install_allowed": false,
    "user_confirmation_required": true,
    "blocks_install": true
  },
  "one_step_action": {
    "action_type": "ask_user_before_install"
  }
}

Agents must execute the structured fields rather than infer policy from prose.

Intelligence States

• strict_reviewed: versioned evidence, technical scan, community-source check, and independent recalculation completed.
• curated_baseline: manually curated fallback information with limited evidence.
• automatic_assessment: automated coverage that cannot authorize automatic installation.
• unknown: no matching intelligence record.

ASL evaluates observable behavior and installation context. It does not label a component malicious without evidence.

The public fallback contains strict reviewed records and curated fallback baselines. Automatic assessments are available through the online service but cannot authorize automatic installation.

Privacy

Online lookup uses public component metadata only. Do not submit:

• private source code or file contents
• API keys, passwords, tokens, or SSH keys
• browser cookies or personal browser history
• private prompts, memories, or customer records
• secrets from environment variables

See PRIVACY.md and SECURITY.md.

Local Fallback

$env:ASL_MODE="local"
npx -y agent-security-lens

Local fallback provides basic rules and a limited public intelligence baseline. The current online intelligence service should be preferred when available.

Verification

npm run verify:public

The scoring method is documented in ASL Agent Component Safety Standard v0.2.

License

Apache-2.0