Your coding agent can ship 10 commits before your security scanner finishes its first scan. That gap is where the vulnerabilities live. Semgrep's new Agentic AppSec Platform, launched November 12, puts a security agent directly inside the coding agent loop — scanning and fixing code the moment it's written, not after it lands in a pull request. A 2026 study of 22,000 developers found a 31.3% rise in pull requests merged without any review. Something had to give.
Shift Left Is Dead
The "shift left" security model was built for human developers working at human speed. AI coding agents have retired it. When Copilot or Claude Code can generate hundreds of lines per minute, traditional SAST tools become a bottleneck, not a safeguard. The Cloud Security Alliance's 2026 research found that 62% of AI-generated code contains design flaws or known vulnerabilities, with CVEs from AI-generated code hitting 74 by March of this year.
Semgrep's answer is to stop treating security as a post-generation step and start running it concurrently with the coding agent itself. Security as a co-agent, not a CI gate.
What Shipped
The Agentic AppSec Platform has three components, each targeting a different stage of the developer workflow.
Guardian is the in-IDE piece. It runs next to Copilot, Codex, and Claude Code — scanning each change as the coding agent writes it, generating fixes automatically before you've staged the file. The fix rate for vulnerabilities caught at generation is 3.6x higher and 38x faster than findings caught at PR review. At launch, Semgrep added GitHub Copilot support, completing the main trio of AI coding tools. Guardian's policies and developer rules are now generally available.
Agentic Security Workflows handle the harder class of bugs: authentication flaws, authorization gaps, injection chains, and business logic vulnerabilities that pattern matching misses entirely. These run as background agents powered by Mandoline, Semgrep's program analysis engine. In Semgrep's benchmarks, the platform finds more than twice as many real vulnerabilities as frontier models alone, while cutting token costs by 65%.
The MCP Server is in public beta and arguably the most immediately useful piece for most developers. It exposes the full Semgrep platform as callable tools inside any MCP client. Connect it to Claude Code, Cursor, Codex, VS Code, or Windsurf, and your agent can triage findings, run scans, generate fixes, and file tickets without leaving the editor.
Connect It in One Line
Setup is minimal. For Claude Code, it is a single command:
claude mcp add semgrep uvx semgrep-mcp
Or connect to the hosted server directly:
claude mcp add --transport http --scope user semgrep https://semgrep.dev/api/mcp
For Cursor, add to ~/.cursor/mcp.json:
{
"semgrep": {
"command": "uvx",
"args": ["semgrep-mcp"]
}
}
Once connected, your agent can query Semgrep for findings, generate fixes for specific vulnerability classes, and push tickets — all within the same context window where the code was written.
Why the Technology Is Different
The token efficiency numbers come from Mandoline, Semgrep's deterministic program analysis engine. Most AI-driven security tools burn tokens asking a model to scan entire codebases; LLMs miss up to 90% of endpoints when working blind. Mandoline runs a taint analysis first — tracking data flows through the code — then hands the model precise, targeted facts. It scored 100% on the OWASP Benchmark for Java and 98.28% on Python with zero AI token usage on those tests.
The cost difference matters: unguided AI security analysis runs $72.55 per true positive. With Mandoline directing the analysis, that drops to $0.62. That is not a margin improvement — it is a category change.
The Broader Signal
Forrester named "Agentic Development Security Tools" a new market category in Q4 2026. Semgrep is positioning as the foundation layer: deterministic analysis feeding AI reasoning, integrated via MCP into every major coding tool. The company will demo the platform at AWS re:Invent (booth #638, Nov 30-Dec 4).
The Agentic AppSec Platform is available now for Semgrep customers. The MCP server is in public beta — no existing Semgrep account required to connect it to Claude Code or Cursor today.

