docs / cli/mcp-server

MCP server

Local-only stdio server that protects your IDE buffer in real time.

Overview

The MCP server runs locally over stdio. It receives your IDE buffer on every keystroke, scans it sub-50ms, and returns findings inline. No network in deterministic mode. No persistence.

The command ekmire mcp start is meant to be launched by your IDE as a subprocess — not run manually in a terminal. The companion command ekmire mcp audit scans all MCP config files on your system for misconfigurations and can be run manually.

Supported IDEs

Cursor, Claude Code, Continue.dev, Zed. Setup is a single line in your IDE's MCP config file.

// Cursor: ~/.cursor/mcp.json
// Claude Code: ~/.claude/settings.json  (under mcpServers)
{
  "mcpServers": {
    "ekmire": { "command": "ekmire", "args": ["mcp", "start"] }
  }
}

See IDE integration for Continue.dev and Zed configs.

Exposed tool: audit_code_buffer

The MCP server exposes one tool: audit_code_buffer. Your IDE calls it automatically; you don't invoke it directly.

PARAMETER	TYPE	REQUIRED	DESCRIPTION
content	string	yes	File content to analyse.
filename	string	no	Filename for context (e.g. `auth.py`). Helps the regex engine match file-type patterns.
mode	string	no	`deterministic` (default) · `ai` · `full`. See modes below.

Modes

deterministic (default) — regex rules + Python AST only. Sub-50ms. No network. Safe for every keystroke.
ai — LLM analysis only. Requires authentication and a configured LLM provider.
full — deterministic first, then LLM. Use for on-save rather than on-keystroke to control token cost.

To enable AI mode, pass --mode=full as an arg in your IDE config:

{
  "mcpServers": {
    "ekmire": { "command": "ekmire", "args": ["mcp", "start", "--mode=full"] }
  }
}

NOTEThe --mode flag sets the default for all calls the IDE makes. The IDE can still override it per-call via the mode parameter in audit_code_buffer.

How the IDE flow works

When you add ekmire to your MCP config, your IDE:

Launches ekmire mcp start as a subprocess on IDE startup.
Sends an MCP initialize handshake — ekmire responds with its protocol version and capabilities.
Calls tools/list — ekmire returns the audit_code_buffer tool schema.
On every buffer change (or on save in full mode), calls tools/call with audit_code_buffer and the current file content.
ekmire returns a JSON array of findings — the IDE renders them as inline annotations.

The process is entirely local. JSON-RPC messages travel over stdin/stdout between the IDE and the ekmire subprocess. Nothing leaves your machine in deterministic mode.

Response shape:

{
  "findings": [
    {
      "severity": "critical",
      "rule_id": "HARDCODED_AWS_KEY",
      "line": 14,
      "description": "Hardcoded AWS Access Key",
      "remediation_prompt": "Move to environment variable or secrets manager."
    }
  ],
  "total": 1,
  "engine": "deterministic",
  "latency_ms": 12
}

MCP ecosystem audit

A separate command that scans all MCP config files on your system for misconfigurations. Run it manually — it doesn't require a running IDE.

$ ekmire mcp audit

It scans ~/.cursor/mcp.json, ~/.claude/settings.json, VS Code workspace configs, and other known MCP config locations. Results are printed to the terminal and posted to the dashboard under Events → Source: mcp.

What it checks per server entry:

Is the server bound to 0.0.0.0? (network exposure)
Does the URL have a pinned version hash? (rug-pull risk)
Are there hidden instructions in tool descriptions? (tool poisoning)

What it catches

MCP tool poisoning, hidden instructions in tool descriptions, MCP servers bound to 0.0.0.0 (network exposure), servers with unpinned/ambiguous command paths, and prompt injection strings embedded in source code before they hit disk.

First-run setup (init)