@essentianlabs/radar-mcp

Early beta for evaluation only.

RADAR MCP is in active evaluation. If you try it, assume it's experimental and tell me where it breaks — technically or conceptually.

Local MCP server that wraps @essentianlabs/radar-lite as a tool Claude can call before executing actions. Runs entirely on your machine — same privacy model as radar-lite.

What It Does

Exposes one tool: radar_assess — assess an intended action for risk before executing it. Returns PROCEED, HOLD, or DENY with Vela Lite's reasoning and strategy options.

When Claude is about to take an action (send an email, delete data, call an API), it calls radar_assess first. If the verdict is HOLD, Claude stops and explains the risk. If the verdict is DENY, the action is blocked by policy.

Everything runs locally. The MCP server calls radar-lite on your machine. radar-lite calls your own LLM provider key (OpenAI, Google, or Anthropic). No calls are made to EssentianLabs servers.

Claude → radar-mcp (local) → radar-lite (local) → your LLM key (OpenAI/Google)

Prerequisites

• Node.js 18+ — download
• npm — comes with Node.js
• An LLM API key — OpenAI (recommended) or Google. Required for Vela Lite T2 assessments. Without a key, radar-lite falls back to rules-engine-only scoring (T1 works, T2 returns HOLD without reasoning).
• Claude Code (VS Code extension) or Claude Desktop — whichever you use as your agent

Install

Step 1: Install packages

npm install @essentianlabs/radar-lite@latest @essentianlabs/radar-mcp@latest

Using @latest ensures version compatibility between the two packages. If you get an ERESOLVE peer dependency error, update radar-lite first: npm install @essentianlabs/radar-lite@latest.

Step 2: Register with Claude Code

npx radar-mcp install

This does two things:

1. Registers the MCP server with Claude Code at user scope (available in all sessions)
2. Adds a RADAR instruction to ~/.claude/CLAUDE.md so Claude automatically calls radar_assess before taking external actions

To verify:

claude mcp list

To uninstall (removes MCP server and CLAUDE.md instruction):

npx radar-mcp uninstall

To update:

npm update @essentianlabs/radar-mcp
npx radar-mcp install

The install command is idempotent — re-running it refreshes the MCP server registration and the CLAUDE.md instruction with any changes from the new version.

Step 4: Configure your LLM provider

Launch the radar-lite dashboard:

npx @essentianlabs/radar-lite dashboard

Go to Settings and configure your LLM provider and API key. Recommended: OpenAI (see Segregation of Duties below).

Or create ~/.radar/.env manually:

LLM_PROVIDER=openai
LLM_API_KEY=sk-your-openai-key
RADAR_AGENT_ID=my-agent

RADAR_AGENT_ID is optional — defaults to claude-code. Set it to identify this agent in the dashboard when running multiple agents.

Supported providers: openai (recommended), google, anthropic (not recommended — see Segregation of Duties below).

Without an LLM key, radar-lite still works but only uses the deterministic rules engine. T1 actions get scored and return PROCEED/HOLD. T2 actions return HOLD without Vela's reasoning or strategy options.

Claude Desktop (alternative)

If you use Claude Desktop instead of Claude Code, add to your claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %LOCALAPPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "radar": {
      "command": "node",
      "args": ["/absolute/path/to/radar-MCP/bin/radar-mcp.js"]
    }
  }
}

Restart Claude Desktop after saving.

Cursor

Add to your Cursor MCP settings (.cursor/mcp.json in your project, or global settings):

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

Add this to your .cursorrules file so Cursor calls RADAR automatically:

Call radar_assess before every task. It must be your first tool call.
If the verdict is PROCEED, continue. If HOLD or DENY, stop and explain the verdict to the user.

Step 5: Verify

Restart any open Claude Code sessions — MCP servers are loaded when a session starts. Existing sessions won't see the radar tool until restarted.

Then ask Claude to do something that would trigger an assessment — for example:

"Send a bulk email to 5,000 subscribers announcing our new feature"

Claude should call radar_assess before acting. You'll see the tool call and the verdict in the conversation.

Note: Dashboard changes (sliders, human review, trigger policies, LLM keys) take effect on the next radar_assess call within any running session — no restart needed for config changes. Only the initial install requires a session restart.

Segregation of Duties

Since Claude is the calling agent, we recommend using a different LLM provider for RADAR assessments — openai or google. T1 uses a fast model for quick scoring. T2 uses a reasoning model for deeper review. Model selection is handled automatically based on your configured provider.

If RADAR_LLM_PROVIDER is set to anthropic, the server logs a warning.

Tool Schema

radar_assess

Input:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | action | string | Yes | What the agent intends to do — be descriptive for accurate scoring | | activityType | enum | Yes | One of: email_single, email_bulk, publish, data_read, data_write, data_delete_single, data_delete_bulk, web_search, external_api_call, system_execute, system_files, financial | | agentId | string | No | Identifier for this agent (used for per-agent config and history) |

Output:

| Field | Type | Description | |-------|------|-------------| | status | "PROCEED" | "HOLD" | "DENY" | Primary verdict | | verdict | string | Same as status (backwards compatibility) | | proceed | boolean | true if safe to proceed | | reviewRequired | boolean | true on HOLD — explicit human/system review needed | | tier | number | Assessment tier (1 or 2) | | riskScore | number | 1–25 risk score | | triggerReason | string | Why this score was assigned | | vela | string | null | Vela Lite's formatted assessment | | options | object | null | T2 strategy options: { avoid, mitigate, transfer, accept } | | recommended | string | null | Recommended strategy | | holdAction | string | Configured action on HOLD (halt, queue, log_only, notify) | | callId | string | Unique assessment ID for this call | | policyDecision | string | Policy outcome (assess, human_required, no_assessment, deny) | | wouldEscalate | boolean | Whether this would be T3/T4 on the full RADAR server | | escalateTier | number | null | Escalation tier (3 or 4) if applicable |

Verdict behaviour

• PROCEED — Below review threshold. Agent may continue.
• HOLD — Requires review. Comes with Vela's reasoning, four strategy options (avoid/mitigate/transfer/accept), and a recommended strategy.
• DENY — Hard stop. Blocked by policy or extreme risk. Can only be overridden via the radar-lite dashboard — not by the agent.

Dashboard

radar-lite includes a local dashboard where you can view assessment history, risk scores, Sankey flow charts, agent stats, and configure activity sliders and policies.

To launch it:

npx @essentianlabs/radar-lite dashboard

This starts the dashboard at http://localhost:4040. All assessments made through the MCP server are automatically recorded in radar-lite's local SQLite database and visible here.

If you get a "not found" error, install radar-lite globally first:

npm install -g @essentianlabs/radar-lite

Then run:

radar-lite dashboard

Dependencies

| Package | Version | What it does | |---------|---------|--------------| | @modelcontextprotocol/sdk | ^1.28.0 | MCP server framework (stdio transport) | | zod | ^3.25.0 | Input schema validation | | @essentianlabs/radar-lite | ^0.3.0 | Risk assessment engine (peer dependency) |

Advisory notice

RADAR produces risk intelligence, not safety assurance. It structures reasoning — it does not validate decisions.

• RADAR assesses the action description supplied by the developer or agent. It does not verify, monitor, or control the real-world action that is actually executed.
• A PROCEED verdict means "not held by this assessment." It is not authorization, approval, certification, legal advice, or safety validation.
• RADAR can produce a PROCEED verdict for actions that later prove harmful, incorrect, unethical, or non-compliant. The assessment reflects what was described, not what occurs.
• Liability remains with the developer, operator, and end user. RADAR does not transfer, reduce, or share liability for actions taken.
• If an external LLM provider is configured, action text leaves the local machine and is sent to that provider under your own account and API terms.
• RADAR cannot enforce decisions in MCP integrations. Claude seeing HOLD or DENY is advisory. The system prompt instruction is what makes Claude respect verdicts. RADAR does not intercept or block actions by itself.

This is a beta release. Not recommended for enterprise or production use without independent legal and compliance review. By installing this package you agree to the Beta Terms of Use.

See the radar-lite README for full advisory terms.

Privacy and Data Flow

• No calls to EssentianLabs servers — everything runs locally
• LLM calls go to your configured provider (OpenAI, Google, or Anthropic) using your API key, at your cost. Action descriptions are sent to the provider for assessment. Review your provider's data retention policies.
• Without an LLM key — no external calls are made. The rules engine runs entirely locally.
• Assessment history — stored locally in SQLite at ~/.radar/register.db. Action hashes only — never action text.
• No telemetry, no analytics, no phoning home

CLI

npx radar-mcp install             # Register MCP server + add CLAUDE.md instruction
npx radar-mcp install --dashboard # Same + open the dashboard
npx radar-mcp uninstall           # Remove MCP server + remove CLAUDE.md instruction
npx radar-mcp --version           # Print version

License

MIT — see LICENSE.

Copyright 2026 EssentianLabs.