Semgrep MCP: security scans before agent code ships

Coding agents write code fast, but they won’t necessarily notice string-built SQL, hardcoded secrets or unsafe deserialization on their own. Semgrep MCP lets the agent call Semgrep, a deterministic static analyzer: after writing or editing code it runs a scan and fixes against concrete rule matches and line numbers. The server now lives in the main Semgrep repository and starts with the semgrep mcp subcommand; the standalone semgrep/mcp repository is archived. The Semgrep repo had about 17k stars as of 2026-10-06.

What it does

• Code scanning: semgrep_scan checks files against Semgrep rules, drawing on a registry of more than 10,000 rules across many languages.
• Custom rules: semgrep_scan_with_custom_rule runs a rule you supply. The built-in write_custom_semgrep_rule prompt helps the model write accurate rules, and semgrep_rule_schema documents the format.
• Dependency risk: semgrep_scan_supply_chain looks for supply chain issues in dependencies.
• Syntax trees: get_abstract_syntax_tree returns the AST so the model can reason about structure or debug a rule.
• Platform findings: with SEMGREP_APP_TOKEN set, semgrep_findings reads existing findings from Semgrep AppSec Platform.

Who it’s for

• Developers who want Claude Code or Cursor to pass a security scan before handing back every change.
• Security teams already on Semgrep who want the same rules applied to AI-generated code.

Setup

Install the Semgrep CLI, then start the MCP server in stdio mode:

brew install semgrep
semgrep mcp

Claude Code and Cursor users can instead install the Semgrep plugin (Semgrep Guardian) from the plugin marketplace and run /setup-semgrep-plugin. With Docker: docker run -i --rm semgrep/semgrep semgrep mcp -t stdio.

Our take

Among code-scanning servers, Semgrep’s edge is determinism: the same code and rules give the same result, unlike asking a model to review code, which varies run to run. The Semgrep Guardian plugin bundles the MCP server with hooks and skills, scans files the agent generates and prompts it to rewrite until the scan is clean or you dismiss the finding. Know the limits: the README says the project is under active development, reading platform findings needs a Semgrep account and token, and community rules cover less than the paid Semgrep Code, Supply Chain and Secrets products. The MCP code ships inside the Semgrep CLI, licensed LGPL-2.1.