Hejdar MCP Server is a runtime policy enforcement tool provided by Hejdar for AI agents, released in the form of an MCP (Model Context Protocol) server. Its core purpose is to let an agent, before executing an action, first initiate a check against organizational policy to determine whether that action is permitted.
Main capabilities: The server exposes hejdar_evaluate as an MCP tool, which any MCP-compatible agent (Claude, ChatGPT, Cursor, custom agents) can call. The tool evaluates whether an agent action complies with the organization's security policy, and the returned decision is one of ALLOW, DENY, or WOULD_DENY, and it can work in observation mode or enforcement mode. The server itself is merely a lightweight wrapper around the Hejdar API (POST /v1/evaluate) and does not contain policy logic; all decisions come from policies configured within the Hejdar organization.
Usage: It can be installed via pip install hejdar-mcp, or run directly with uvx hejdar-mcp. Before use, you need to register an account at app.hejdar.com and create an API key in Settings → API Keys. Then configure the MCP client: for Claude Desktop, you can add the configuration in ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows; it can also be added to the MCP settings of Claude Code. The documentation also recommends adding the relevant system prompt patterns to the agent's system prompt for best results.
Security and limitations: The API key is read only from environment variables and will not be hardcoded or exposed in tool input or output; all input is validated and sanitized before being forwarded to the API; error responses do not leak internal details, API keys, or stack traces; all API calls are forced to use TLS. Note that this MCP server does not contain policy logic, and actual decisions depend on the policies configured in the Hejdar organization, so before use you must have an Hejdar account and API key. The documentation does not provide specific pricing, charging model, or available platform details; this information is not currently available.
