Argus

What alexar76/argus MCP server does

The alexar76/argus MCP server gives an MCP client access to three operations: submit a bounded natural-language task, inspect service status, and discover available capabilities. Its primary task interface is argus_ask, which accepts a task and can also receive a response format or focus. For example, an agent can request a short, bullet-based summary of a URL.

argus_status returns JSON suitable for a liveness check before starting heavier work. argus_capabilities returns a JSON tool catalog and can optionally include schemas. These operations let a client inspect the service without consuming model tokens for an ordinary task.

How it works

The server uses stdio transport through the argus mcp entrypoint and is built with the official TypeScript MCP SDK. It is intended for MCP hosts that support local stdio processes, including Claude Desktop and Cursor.

Before a third-party MCP server is allowed to run, WARDEN applies a sequence of checks. The documented sequence includes static tool-definition scanning, a threat-feed check, origin validation, and definition pinning to detect changes. The checks use data already held locally, so WARDEN does not require network access to produce a verdict.

The agent core uses a bounded reasoning budget and reports failures or budget stops through argus_ask. Wallet, crypto, paid invocation, and on-chain settlement features are not required for the MCP server’s basic operation and are off by default.

Setup and configuration

Install the scoped npm package globally, then start the MCP entrypoint:

npm install -g @alexar76/argus3
argus mcp

A host configuration uses the installed argus executable with mcp as its argument:

{
  "mcpServers": {
    "argus": {
      "command": "argus",
      "args": ["mcp"]
    }
  }
}

The repository also documents a source-based route using npm install, npm run build, and node dist/index.js mcp. Crypto can be enabled with ARGUS_CRYPTO_ENABLED=1 together with a wallet, but that setting is optional and is not needed for the standard MCP use case.

Tools and capabilities

• argus_ask: Runs bounded natural-language work through the agent core. It supports optional response-format and focus parameters and returns plain text, or an MCP error when the task fails or reaches its budget limit.
• argus_status: Returns JSON service status. Use it for basic or full liveness checks.
• argus_capabilities: Returns the available tool catalog, with an option to include schemas.
• WARDEN protection: Evaluates third-party MCP servers before tool execution using static scanning, threat data, origin checks, and definition pinning.
• Model-provider flexibility: The broader ARGUS agent supports Anthropic-native, OpenAI-compatible, and local Ollama providers, according to the project documentation.

Limitations and notes

The alexar76/argus MCP server is a local stdio process rather than a hosted MCP endpoint. Its documented compatible hosts are Claude Desktop, Cursor, Glama, and other stdio-capable MCP clients; this listing only identifies Claude Desktop and Cursor as supported client slugs.

The repository is a read-only mirror of a canonical monorepo, and pull requests to the mirror are not accepted. WARDEN’s protection applies to third-party MCP servers that ARGUS evaluates; it does not turn every task into a general-purpose security scan. Crypto and wallet functions are opt-in, so users seeking on-chain settlement must configure those features separately. The project is licensed under MIT.