Marshell is a direct participant in the AI agent ecosystem, specifically focused on the "agent skills" and security layer. They provide a standardized way for coding agents to interact with security databases, effectively giving an LLM-based agent the 'eyes' to see vulnerabilities and the 'instructions' to fix them. Their product integrates with prominent agents like Claude Code and Cursor, making them a key utility in the developer agent stack.
By championing the use of AGENTS.md and custom skills, Marshell is helping define how third-party tools can extend the capabilities of coding agents. They are active in the security/supply chain portion of the stack, ensuring that the speed of agentic code generation does not result in a proliferation of vulnerable software. Their presence is significant for anyone building autonomous coding workflows where human security oversight is a bottleneck.
Marshell is a San Francisco-based security firm building software specifically for the AI-coding era. Founded in April 2026 by Arsenii Trutnev and Danil Sokolov, the company focuses on the security risks introduced by the rapid adoption of AI coding agents. As developers increasingly use tools like Cursor, Claude Code, and Windsurf to generate software, the volume of third-party dependencies pulled into projects has increased. Marshell operates on the premise that while agents are excellent at writing code, they are not inherently programmed to audit the security of the packages they install.
The company’s first product, Chifu, is a security tool designed to bridge this gap. It consists of a Command Line Interface (CLI) and an "agent skill" that integrates directly with various coding environments. When a developer or an agent runs a scan, Chifu matches every dependency — both direct and transitive — against a live corpus of over 65,000 vulnerabilities sourced from the National Vulnerability Database (NVD) and Open Source Vulnerability (OSV) database.
What makes Marshell different from traditional security scanners is the intended workflow. Traditional tools often produce a report of "flags" that a human developer must then interpret and fix. Marshell is built for a loop where the agent is the primary actor. Once the dep-guard skill is installed, coding agents can run security checks autonomously before a commit is made. If a vulnerability is found, the agent is instructed to read the findings, upgrade the package, and resolve any breaking changes caused by the update. This process continues until the environment is clean.
The founders bring a background in offensive security and infrastructure to the project. Arsenii Trutnev, the CEO, began hunting bugs on web platforms at age 14 and started Marshell after experiencing firsthand how a single unpatched vulnerability could impact a production system. Danil Sokolov, the CTO, is a hackathon specialist who manages the infrastructure required to keep the vulnerability data refreshed and the scanning process fast.
Marshell currently supports projects using package.json and lockfiles from npm, pnpm, yarn, and bun. Its integration strategy is broad, targeting any agent that can read an AGENTS.md file. This includes Claude Code, Cursor, Codex, OpenCode, Windsurf, Gemini CLI, and Cline.
The company's long-term roadmap involves moving beyond simple dependency scanning. Marshell intends to deploy specialized AI security agents for different surfaces, including web applications, APIs, infrastructure, and mobile development. The goal is to provide a comprehensive security layer that AI-first startups can adopt as easily as they adopt payment or hosting infrastructure. The business operates on a freemium model, offering a baseline of free scans per month with paid tiers for higher volume and continuous monitoring.
A CLI and agent skill that catches vulnerable dependencies.
Marshell is hiring.