---
title: Quickstart
description: Run secure-mcp 2.x from repository inventory through stack-aware checks, evidence verification, and a structured remediation report for an authorized code review.
sidebar:
  label: Quickstart
  order: 3
---

## Quickstart

Install first: clone, then run the one-command bootstrap `./scripts/setup.sh` (or `.\scripts\setup.ps1` on Windows). It builds the server, prompts for the filesystem allowlist when it is unset, wires the skill and MCP server into pi, Cursor, and Codex, and verifies the result — then restart your agent. Prefer an allowlist parent that contains the repositories you review. The server requires MCP protocol `2026-07-28`. Grok Build TUI is unsupported until it negotiates that revision.

The safest useful audit is staged. Each phase leaves a small artifact the next phase can use, so agents do not spend context on irrelevant checklists or jump from a regex hit to a confident claim.

1. **Inventory the target**

    Call `secure_mcp_list_project_structure` with an absolute `project_root`. Record the file count, sample paths, and any scope limits.

2. **Profile the architecture**

    Call `secure_mcp_analyze_architecture` with `stack: "auto"`. Keep its detected stacks, trust boundaries, `recommended_packs`, and `pack_batches`.

3. **Load only the relevant knowledge**

    Pass the first `pack_batches` entry to `secure_mcp_get_knowledge_pack` with the default `detail: "summary"`. Load later batches only when the review needs them.

4. **Run category checks**

    Run authentication, injection-risk, and secret-hygiene checks in parallel when the client supports it. Add the remediation threat model when control placement needs more structure.

5. **Confirm the candidates**

    Open each high- or critical-confidence candidate at its cited file and line. Verify the data flow, remove fixtures that are out of scope, and merge duplicates.

6. **Produce the report**

    Send the confirmed findings to `secure_mcp_produce_findings` with `dedupe: true`, a minimum severity/confidence, and `response_format: "markdown"` for the human-facing handoff.

## A focused architecture call

```json
{
  "project_root": "/absolute/path/to/target",
  "stack": "auto",
  "max_files": 400,
  "focus_paths": ["src", "app"]
}
```

Use `focus_paths` for a drill-down after the first repository-wide pass. Keep the initial `max_files` bounded; the server already applies depth, byte, and ignore-pattern caps.

## The non-negotiable rule

> Treat heuristic hits as candidates, not confirmed vulnerabilities. Read the cited code and verify the data flow before reporting impact.

For the full sequence, intermediate artifacts, and defensive boundaries, read [Agent workflow](/docs/agent-workflow).