---
title: secure-mcp
description: "Defensive, agent-first security audits over strict MCP v2 (2026-07-28): scope repositories, route stack-aware guidance, and produce structured remediation findings."
sidebar:
  label: Start here
  order: 1
---

<Badge variant="success">read-only by design</Badge> <Badge variant="accent">stdio transport</Badge>

## Security review that agents can carry forward

`secure-mcp` gives coding agents a small, structured toolkit for defensive code review. It inventories a repository, detects its stack, loads only the knowledge packs it needs, and turns likely weaknesses into remediation-ready findings. Version 2.x requires MCP protocol revision `2026-07-28`; legacy 2025-era handshakes are rejected.

The server is intentionally local. It reads files under an approved `project_root`, never executes the target project, and keeps stdout reserved for MCP JSON-RPC.

<CardGroup cols={2}>
  <Card title="Clone and install" href="/docs/installation" icon="rocket">
    Clone, build, and install the server plus agent workflow for pi, Cursor, and Codex, or connect another strict v2 stdio client.
  </Card>
  <Card title="Client compatibility" href="/docs/clients" icon="plug">
    Verified configuration shapes for Codex, Cursor, Claude Desktop, Claude Code, VS Code / Copilot, pi, and generic stdio clients.
  </Card>
  <Card title="Run the audit workflow" href="/docs/quickstart" icon="workflow">
    Follow the multi-phase sequence from inventory to confirmed findings and a human-facing remediation plan.
  </Card>
  <Card title="Explore the tool surface" href="/docs/tools" icon="scan-search">
    See what each tool reads, returns, and contributes to the final report.
  </Card>
  <Card title="Host these docs" href="/docs/hosting" icon="github">
    Build and verify the Blume site in GitHub Actions, then publish it through a separate Cloudflare Pages or Workers configuration.
  </Card>
</CardGroup>

## The operating model

<Columns cols={3}>
  <Column>
    ### Scope first

    Start with project structure and architecture so scans stay bounded and stack-aware.
  </Column>
  <Column>
    ### Remediate clearly

    Every finding carries evidence, classification, impact, remediation, residual risk, and verification.
  </Column>
  <Column>
    ### Verify before reporting

    Heuristics are candidates. Agents confirm data flow before treating a signal as a finding.
  </Column>
</Columns>

:::note
This project is for authorized defensive review only. It does not generate exploits, execute code from the target repository, or use discovered secrets against systems.
:::

## What it covers today

| Surface | Coverage |
| --- | --- |
| Web | TypeScript, Next.js App Router, API handlers, cookies, CSRF, and `NEXT_PUBLIC_` exposure |
| Apple platforms | Swift, SwiftUI, Keychain, ATS, WebView bridges, deep links, and macOS controls |
| React Native | Expo config, SecureStore, OTA boundaries, deep links, and token storage |
| Cross-cutting | Authentication, injection risks, secrets, threat modeling, and report rollup |

<Card title="Read the design notes" href="/docs/architecture" icon="book-open">
  Understand the process boundaries, filesystem policy, progressive knowledge packs, and extension points.
</Card>