Agentseal

AgentSeal is a security scanner for AI agents that red teams system prompts, audits Model Context Protocol servers and inspects skill and instruction files from a single command line tool. It installs from PyPI or npm and runs four core commands: scan, which fires an adversarial probe suite at a prompt; guard, which audits the agents configured on a machine; watch, which flags changes in real time; and scan mcp, which connects to live MCP servers and tests them at runtime. It is built for developers and security teams shipping agents who want to find prompt extraction, injection, poisoned skill files and risky MCP configurations before an attacker does. Probes cover extraction, injection, multimodal and behavioral attacks, and the scanner checks Cursor rules, skill files and combined skill plus MCP attack paths. AgentSeal also maintains a security analyzed MCP registry in which servers are sandboxed, probed and scored. It works with twelve model providers including Ollama, Claude and OpenAI on a bring your own key basis, so prompts stay with the chosen provider. For pipelines it drops into GitHub Actions, emits SARIF and JUnit output, supports policy as code in a YAML file, and can block pull requests on critical findings. The CLI is published under the Functional Source License. The free tier includes 30 basic probes, the MCP registry, Guard and Watch. Pro unlocks the full probe library, runtime MCP scanning with OAuth, PDF export and the dashboard; Pro pricing is not published and is quoted by the AgentSeal team on request. Because the tool is bring your own key, the only ongoing cost on the free tier is the user's own model usage, which the vendor notes can be nothing when running locally with Ollama.