650,807 tools. Updated 2026-10-14 20:36
"Security vulnerabilities in MCP Server related to subprocess execution" matching MCP tools:
- Get Lenny Zeltser's CTI cross-server handoff routes — when this MCP server can't fulfill a request, which other MCP servers (or fallback workflows) to consult. Surfaces a compact subset of `cti_load_context`. This server never requests your campaign or threat-intel notes and instructs your AI to keep them local—templates and guidelines flow to your AI for local analysis.ConnectorNo auth
- Connectivity check — returns server version and current timestamp. Use to verify MCP server is reachable before calling other tools.ConnectorNo auth
- Connectivity check that confirms the Nordic MCP server process is responding. Use this at the start of a session to verify the server is reachable before making other calls. Do not use as a proxy for database health — the server can respond while the Qdrant vector database is temporarily unavailable. To confirm data availability, call search_filings directly. Returns: A greeting string: "Hello {name}! Nordic MCP server is running."ConnectorNo auth
- Get a security advisory (vulnerability) by its key. Returns a security advisory by key, for example a GHSA id taken from a version's advisoryKeys, including the title, CVE aliases, CVSS v3 score and vector, and a link to the full record on osv.dev. Use this only when you already have an advisory ID from get_package_version's advisoryKeys. There is no search here. To find out whether a version has vulnerabilities at all, call get_package_version first; this tool explains one advisory in depth.ConnectorNo auth
- No arguments. Returns how many MCP servers have been read at source level, and the share of them with each category of finding (credential access, network egress, install-time execution, prompt-injection surface). Use this to judge whether checking a specific server is worth it before you look one up. It reports aggregate counts only - no per-server findings, and no verdict about any individual server.ConnectorNo auth
- Start an OAuth Connect and get the consent URL to open in a browser. Pass provider (github|google|x|atlassian|microsoft|slack|notion|gitlab|discord|reddit|linear — account-wide, reused everywhere) OR server_url (any MCP server; Rokha runs the MCP auth-spec handshake, grant bound to that server). On approval the token lands in the vault under the returned alias. Requires a logged-in identity.ConnectorNo auth
Matching MCP Servers

execution-run-mcpofficial
AlicenseAqualityNot gradedmaintenanceEnables AI agents to interact with the Execute.run bot API for managing Shell balances, transferring funds, and executing LLM requests. It provides tools for identity verification, transaction tracking, and performing compute tasks through the Execute.run platform.6-- FlicenseBqualityDmaintenanceA Model Context Protocol server that enables LLM applications to safely execute shell commands with error handling and timeout settings.1-
Matching MCP Connectors
Free read-only Base Mainnet transaction Preflight and Verification; no signing or custody.
- execution-market-mcpOAuth
Execution Market is the Universal Execution Layer — infrastructure that converts AI intent into physical action. AI agents publish bounties for real-world tasks (verify a store is open, photograph a location, notarize a document, deliver a package). Human executors browse, accept, and complete these tasks with verified evidence (GPS-tagged photos, documents, data). Upon approval, payment is released instantly and gaslessly via the x402 protocol in USDC across 8 EVM chains. Key cap
- Get the Sigistry MCP App safety criteria to audit an MCP App (MCP Apps / SEP-1865: a tool that returns an interactive ui:// HTML resource the host renders in a sandboxed iframe) BEFORE shipping it. Returns two groups of checks: spec conformance (ui:// scheme, the text/html;profile=mcp-app mimeType, tool-to-UI linkage via _meta.ui.resourceUri, a text fallback, visibility values, and CSP coverage of external origins) and security hygiene the iframe sandbox does not cover (unsafe DOM sinks, embedded secrets, host-message origin handling, remote scripts). Call this when the user is building or reviewing an MCP App; then apply each check to the app source you have in context and report findings. Nothing is sent to this server; an interactive browser version is at https://sigistry.com/mcp-app-checker.ConnectorNo auth
- Get the Sigistry MCP App safety criteria to audit an MCP App (MCP Apps / SEP-1865: a tool that returns an interactive ui:// HTML resource the host renders in a sandboxed iframe) BEFORE shipping it. Returns two groups of checks: spec conformance (ui:// scheme, the text/html;profile=mcp-app mimeType, tool-to-UI linkage via _meta.ui.resourceUri, a text fallback, visibility values, and CSP coverage of external origins) and security hygiene the iframe sandbox does not cover (unsafe DOM sinks, embedded secrets, host-message origin handling, remote scripts). Call this when the user is building or reviewing an MCP App; then apply each check to the app source you have in context and report findings. Nothing is sent to this server; an interactive browser version is at https://sigistry.com/mcp-app-checker.ConnectorNo auth
- Scan source code for injection vulnerabilities: SQL injection, command injection, path traversal via unsafe string concatenation/unsanitized input. Supports Python, JavaScript, TypeScript, Java, Go, Ruby, Shell, Bash. Use to detect input-handling bugs; for secrets use check_secrets. Companion code-security tools: check_secrets (hard-coded credential detection), check_dependencies (known-CVE vulnerability audit), check_headers (live HTTP security-header validation), scan_headers (live HTTP scan via domain). Free: 30/hr, Pro: 500/hr. Returns {total, by_severity, findings}. No data stored.ConnectorNo auth
- Get Lenny Zeltser's IR cross-server handoff routes — when this MCP server can't fulfill a request, which other MCP servers (or fallback workflows) to consult. Surfaces a compact subset of `ir_load_context`. This server never requests your incident notes and instructs your AI to keep them local—guidelines flow to your AI for local analysis.ConnectorNo auth
- Get Lenny Zeltser's Malware cross-server handoff routes — when this MCP server can't fulfill a request, which other MCP servers (or fallback workflows) to consult. Surfaces a compact subset of `malware_load_context`. This server never requests your sample, analysis notes, or indicators and instructs your AI to keep them local—guidelines and the report template flow to your AI for local analysis.ConnectorNo auth
- Pre-flight security verdict for an MCP server invocation. Judges BOTH server-level reputation AND the server's dependency graph (npm/pypi) against the DugganUSA threat-intel corpus (1.13M+ IOCs, Shai-Hulud + typosquat + LOLBin families). Returns BLOCK / ADVISORY / REVIEW / ALLOW with severity, evidence, dep-graph summary, and HMAC-signed response. REVIEW means we hold NO RECORD of this server -- not that it is safe. Treat REVIEW as do-not-proceed-blindly: a brand-new attacker-published server looks exactly like this. ALLOW is only returned when we actually resolved the server and scanned its dependency graph; check known_to_us and dep_graph.scanned to confirm. Use this BEFORE invoking any other MCP server tool, especially ones installed from outside the official MCP Registry.ConnectorNo auth
- Free single-call bootstrap aggregator. Returns a compact markdown-like summary with server version, total tools count, tools by category, pricing tiers summary, premium plugins/capabilities count, bundled skills list, and nextAction guidance. Pass agentKnownVersion with the version you currently know to get skillsUpdateRequired + skillsContents (full SKILL.md inline) when the server version differs — this lets you auto-update your local skills in 1 call without a separate sap_skills_bundle. Use this instead of calling sap_agent_start, sap_pricing_catalog, sap_premium_plugin_catalog, sap_skills_list, and sap_get_tool_category_summary separately to reduce bootstrap from 5+ tool calls to 1. SAP MCP execution guidance: Intent: SAP MCP tool workflow. Pricing: free; call directly without x402. Routing: free hosted call; call directly and keep it small/exact when possible. Signer boundary: hosted reads/builders never receive keypair bytes; value-moving results must be finalized locally when signing is required.ConnectorNo auth
- Self-test the MCP server: reports whether the DB pool and execution receiver are reachable. Use to distinguish a backend/infra problem from a bad-input error. Returns: {status: healthy|degraded, checks: {database, receiver}}.ConnectorOAuth
- Search NVD for CVE vulnerabilities by product or component name. Returns CVE ID, description, severity, and CVSS score. Search terms are matched against CVE description text and EVERY word must appear, so pass the product name ("OpenSSL", "log4j", "nginx") optionally with a technical term ("buffer overflow") — not a plain-English question. Use when researching security threats or checking if a known vulnerability affects your systems.ConnectorNo auth
- Return a list of hands-on SecDim Play secure coding challenges (labs) related to a detected or suspected vulnerability. SecDim Play challenges are scored, hands-on labs: find and fix a real vulnerability in running code to earn points and badges. Use this tool to: - Find hands-on SecDim Play labs for specific vulnerabilities like XSS, SQL Injection, etc. - Explore OWASP Top 10 vulnerabilities and related labs - Provide additional resources and guides to help developers improve their secure coding skills For structured tutorial content (text, video, and lab-based courses) on the same vulnerability, use search_learn_courses (SecDim Learn) instead or in addition. Args: search: Search term for the vulnerability (e.g., 'xss', 'sql-injection', 'injection') cwe: Common Weakness Enumeration (CWE) ID to filter by owasp: OWASP category to filter by (e.g., 'a03:2021') technology: Technology or framework to filter by (e.g., 'react', 'django') language: Programming language to filter by (e.g., 'javascript', 'python') difficulty: Difficulty level to filter by (e.g., 'trivial', 'easy', 'medium', 'hard') type: Challenge format to filter by (e.g., 'battle', 'exploitation', 'incident-response') mitre: MITRE ATT&CK ID to filter by (e.g., 'T1102.003') SecDim Play challenges (labs) each simulate a real vulnerability. They are scored according to the following difficulty levels: - Trivial: Easy to find and path vulnerabilities. It can be completed in 5-10 minutes. 1-15 points. - Easy: Known vulnerabilities. It can be completed in 10-30 minutes. 16-35 points. - Medium: Known vulnerabilities but require defence-in-depth patch. It can be completed in 20-30 minutes. 36-70 points. - Hard: Hard to find or patch vulnerabilities. It can be completed in 30-60 minutes. 71-100 points. - Battle: SecDim Flagship attack and defence challenge that require both vulnerability exploitation and mitigation skills. Points are accumulated. Returns: Dictionary containing SecDim Play labs results or error If there are no results, user can perform a manual search on the SecDim Play frontend (SECDIM_PLAY_FRONTEND_BASE_URL)ConnectorNo auth
- Dependency approval decision for coding agents: decide whether an npm dependency or package is safe to install or adopt and return GO, WARN, or BLOCK using vulnerabilities, security, license, maintenance, deprecation, and package metadata. Price: $0.25 via x402. Generate one _salt19_operation_id per intended purchase, preserve it across retries, and add _x402_payment_signature after satisfying the challenge. A stable _salt19_operation_id is required and makes semantic retries at-most-once. _salt19_operation_id is recommended but optional for standard x402 compatibility.ConnectorNo auth
- Runs a free, read-only security scan of a remote MCP server before you connect it. Performs an ordinary MCP handshake (never anything more invasive), lists the server's tools, rates each one's capabilities (private-data access, untrusted content, network egress, destructive actions), flags prompt-injection patterns in tool descriptions, and reports whether the lethal trifecta is reachable. Scan only servers you run or are entitled to connect to.ConnectorNo auth
- List the public MCP servers Sigistry has independently graded against its scorecard rubric (transport, stateless core, lifecycle, authorization, tool design, security hygiene), each with an overall grade A-F and the exact commit graded. Useful before connecting an agent to a third-party MCP server: check whether it has been assessed and how it scored. Use get_scorecard for the full per-axis breakdown.ConnectorNo auth
- List the public MCP servers Sigistry has independently graded against its scorecard rubric (transport, stateless core, lifecycle, authorization, tool design, security hygiene), each with an overall grade A-F and the exact commit graded. Useful before connecting an agent to a third-party MCP server: check whether it has been assessed and how it scored. Use get_scorecard for the full per-axis breakdown.ConnectorNo auth
- Known security vulnerabilities for a package version (npm, PyPI, crates, Go, Maven, NuGet, RubyGems, Packagist) with severity, CVE/GHSA ids, and the version that fixes each. Queries OSV.dev, the open vulnerability database that aggregates GitHub advisories, PyPA, RustSec, Go and more. Pass up to 50 packages at once as "ecosystem:name@version" to audit a lockfile. When to use: For CVEs in one package version; for general package facts use package_info (on the MAD Synapse · Web & Research server, https://agent.maddegen.art/hub/mcp/web). Price: $0.002 per call (10 free/day; after that a payment-required result lists x402 options). Errors: returns isError with a message for invalid input or an upstream failure (not charged).ConnectorNo auth