SAST MCP Server

SAST MCP Server

[PyPI version] [sast-mcp-server MCP server] [Python 3.10+] [License: MIT] [CI] [codecov]

Static Application Security Testing (SAST) for AI agents. A production-ready MCP server that gives any AI agent the ability to scan code for security vulnerabilities.

Supports 11 industry-standard scanners:

Works with any MCP-compatible agent: Gemini CLI, Claude Desktop, OpenAI Agents, Cursor, Windsurf, and more.

Features

• 🔍 11 SAST/SCA/DAST scanners with a unified output format
• 🌳 AST-aware context — shows the full enclosing function, not just a line number
• 📊 Severity & confidence filtering — focus on what matters
• 🔀 Git diff mode — scan only modified files for incremental reviews
• 🙈 Ignore management — suppress false positives with audit trail
• 📄 Pagination — handle large codebases without overwhelming the agent
• 🌐 Dual transport — stdio (local) or Streamable HTTP (remote deployments)
• 🔐 JWT & API key authentication — secure remote deployments
• 📦 One command install — pip install sast-mcp-server
• 🚀 Multi-scanner mode — run all installed scanners in parallel with deduplication
• 📋 SARIF export — CI/CD integration with GitHub, GitLab, Azure DevOps
• 🏗️ IaC scanning — Terraform, Kubernetes, Docker security policies
• 🔑 Secret detection — find hardcoded API keys, tokens, and passwords in code and git history
• 📦 SCA / dependency CVEs — scan lock files for known vulnerabilities against the OSV database
• 🕷️ DAST — dynamic baseline scans of running apps via OWASP ZAP + Docker
• 📈 Baselines & trend tracking — cache scans and diff against a saved baseline
• 🤖 MCP Prompts & Resources — pre-built security workflows and live dashboards for agents
• 📤 Dashboard integrations — push SARIF results to DefectDojo or GitHub Code Scanning
• 🩹 AI-assisted remediation — generate fix prompts and apply agent-written patches via git apply

Quick Start

The server is only as useful as the scanners installed alongside it. Pick the install path that matches how much of the toolset you want out of the box.

Option 1 — Full container (recommended: 9 scanners, zero setup)

docker pull ghcr.io/skyrxin/sast-mcp-server:full

Bundles bandit, njsscan, bearer, semgrep, trivy, checkov, gitleaks, osv-scanner, and grype so scan_all works immediately. Or bring up an HTTP server with one command:

docker compose up          # serves http://localhost:8080/mcp + /health /ready /metrics

Option 2 — pip extra (4 pip-installable scanners)

pip install "sast-mcp-server[scanners]"   # adds bandit, njsscan, semgrep, checkov

Option 3 — minimal / custom

pip install sast-mcp-server               # server only — bring your own scannersuvx sast-mcp-server                        # run without installing

Then install whichever scanners you need (binary scanners aren't pip packages):

pip install bandit njsscan semgrep checkov     # pip-installable# trivy:        https://aquasecurity.github.io/trivy/latest/getting-started/installation/# grype:        https://github.com/anchore/grype#installation# gitleaks:     https://github.com/gitleaks/gitleaks#installing# osv-scanner:  https://google.github.io/osv-scanner/installation/# bearer:       https://docs.bearer.com/installation/# codeql:       https://github.com/github/codeql-cli-binaries/releases

What ships where

At startup the server logs how many scanners it can actually see (e.g. Scanners available: 9/11 (...)), and the list_scanners tool / /ready endpoint report the same — so it's always obvious what you have.

[sast-mcp-server MCP server]

Usage with AI Agents

Gemini CLI

Install as an extension:

gemini extensions install https://github.com/Skyrxin/sast-mcp-server

Or add to your ~/.gemini/settings.json:

{  "mcpServers": {    "sast": {      "command": "uvx",      "args": ["sast-mcp-server"]    }  }}

Claude Desktop

Add to your claude_desktop_config.json:

{  "mcpServers": {    "sast": {      "command": "uvx",      "args": ["sast-mcp-server"]    }  }}

See full Claude Desktop guide.

Cursor IDE

Add to Cursor Settings → MCP Servers:

{  "mcpServers": {    "sast": {      "command": "uvx",      "args": ["sast-mcp-server"]    }  }}

See full Cursor guide.

OpenAI Agents SDK

from agents.mcp import MCPServerStdio
sast_server = MCPServerStdio(command="uvx", args=["sast-mcp-server"])

See full OpenAI guide.

Available MCP Tools

scan_vulnerabilities

Scan a directory for security vulnerabilities using a specific scanner.

scan_all

Run ALL installed scanners in parallel with automatic deduplication. Recommended for comprehensive security scanning.

scan_git_history

Scan the entire .git history for leaked secrets and credentials using Gitleaks.

run_active_scan

Run a dynamic (DAST) baseline scan with OWASP ZAP by orchestrating a Docker Compose stack.

export_sarif

Export scan results in SARIF 2.1.0 format for CI/CD integration.

list_scanners

List available scanners, their installation status, and supported languages.

ignore_vulnerability

Suppress a finding from future scans (with audit trail).

unignore_vulnerability

Re-enable a previously suppressed finding.

list_ignored_vulnerabilities

Show all currently suppressed findings for a project.

save_baseline

Run a scan and cache the results as a named baseline for future trend comparison.

compare_baseline

Compare a fresh scan against a saved baseline to highlight new and fixed findings.

upload_to_defectdojo

Import a SARIF export into a DefectDojo engagement. Requires DEFECTDOJO_URL and DEFECTDOJO_API_KEY environment variables.

upload_to_github

Upload a SARIF report to GitHub Code Scanning. Requires a GITHUB_TOKEN with security_events: write scope.

generate_fix_prompt

Package a cached finding's vulnerable code and context into an LLM-ready prompt that asks for a strict unified diff fix.

apply_patch

Apply an agent-generated unified diff to disk via git apply (paths that escape the target directory are rejected).

evaluate_policy

Run all scanners and return an explicit PASS/FAIL verdict for CI gating.

export_sbom

Run all scanners and export an SBOM / vulnerability report. In CycloneDX mode, if Syft is installed the component inventory is the full dependency list (not just vulnerable packages).

generate_report

Run all scanners and render an executive security report.

compliance_report

Map findings to a compliance framework and report the posture.

scan_image

Scan a container image reference for vulnerabilities and secrets (Trivy or Grype).

remediate_and_verify

Closed-loop remediation: dry-run a patch, apply it, re-scan, and confirm the finding is gone (rolling the patch back on failure).

import_sarif

Ingest an external SARIF file (Snyk, Veracode, CI jobs, …) into the normalized finding pipeline so it joins dedup / baselines / dashboards.

triage_finding

Get an exploitability/false-positive assessment prompt, or record a CycloneDX VEX decision (suppressing dispositions also add the finding to the ignore-list).

comment_on_pr

Post a security summary on a GitHub PR or GitLab merge request. Requires GITHUB_TOKEN or GITLAB_TOKEN (+ optional GITLAB_URL).

notify_slack / notify_teams

Send a notification to a Slack or Microsoft Teams incoming webhook (SLACK_WEBHOOK_URL / TEAMS_WEBHOOK_URL).

create_jira_issue

Open a Jira issue for a finding. Requires JIRA_URL, JIRA_EMAIL, JIRA_API_TOKEN.

Tip: scan_vulnerabilities, scan_all, and scan_git_history accept output_format="json" for machine-readable results in CI/agent pipelines.

Auth: On HTTP transports every tool is scope-gated — scan:read (scans, reports, exports), scan:write (baselines, patches, uploads, notifications), config:write (ignore list). Set SAST_MCP_JWT_SECRET and issue scoped JWTs.

SARIF / CI/CD Integration

Export scan results in SARIF 2.1.0 format for integration with CI/CD platforms:

# In your CI pipeline, use the MCP tool:# export_sarif(target_path=".", scanner_name="semgrep", output_path="results.sarif")
# Then upload to GitHub Code Scanning:# gh api /repos/{owner}/{repo}/code-scanning/sarifs -f [email protected]

Compatible with: GitHub Code Scanning, GitLab SAST, Azure DevOps, VS Code SARIF Viewer.

Remote Deployment (Streamable HTTP)

For remote or cloud-hosted deployments, the 2026 MCP standard uses Streamable HTTP.

You can secure the server with JWT Bearer Authentication by setting a secret. Alternatively, for backward compatibility, you can use a static API key.

# Set JWT secret for secure authenticationexport SAST_MCP_JWT_SECRET="your_hmac_sha256_secret"
# Or use the legacy API key methodexport SAST_MCP_API_KEY="your_secure_api_key_here"
# Start the server with streamable-http transportuv run sast-mcp-server --transport streamable-http --port 8080 --host 0.0.0.0

Note: The old sse transport is deprecated. Please migrate to streamable-http.

Core Workflows

1. Unified Vulnerability Scanning

Run any of the installed scanners individually (scan_vulnerabilities(scanner_name="bandit")) or run all of them at once using scan_all.

2. Deep Secret & Dynamic Scanning

Use scan_git_history to find API keys leaked years ago, or run_active_scan to spin up your application with Docker Compose and test it dynamically with OWASP ZAP.

3. Baseline & Trend Tracking

Save a scan as a named baseline and compare future scans against it to track new vulnerabilities, fixed issues, and severity trends over time.

• save_baseline(target_path=".", tag="main")
• compare_baseline(target_path=".", tag="main")

4. CI/CD Integration

Export scan results to SARIF format (export_sarif) to integrate with GitHub Code Scanning, GitLab SAST, or any other SARIF-compatible platform.

5. MCP Prompts (Security Workflows)

Pre-built security workflows that guide AI agents:

• security_review: Full codebase assessment
• fix_vulnerability: Focused remediation advisor
• pr_security_check: Scan only git diffs and enforce a severity gate
• compliance_report: Generate an OWASP Top 10 or PCI-DSS report

6. MCP Resources (Security Dashboards)

Read-only contextual data for AI agents without running a full scan:

• sast://dashboard/{path}: Security posture dashboard
• sast://config: Server configuration and status
• sast://scanners: Available scanners and languages
• sast://cache/{path}/latest: Latest scan results metadata

7. Dashboard Upload & AI-Assisted Remediation

Push SARIF results to external platforms and remediate findings with agent-written patches:

• upload_to_defectdojo / upload_to_github: Push a SARIF export to a dashboard
• generate_fix_prompt: Build an LLM-ready prompt for a specific finding
• apply_patch: Apply the resulting unified diff via git apply

Docker

Pre-built images are published to GHCR on every release:

docker pull ghcr.io/skyrxin/sast-mcp-server:full       # 9 scanners (recommended)docker pull ghcr.io/skyrxin/sast-mcp-server:minimal    # bandit, njsscan, bearer
# Run as an HTTP serverdocker run -p 8080:8080 -e SAST_MCP_JWT_SECRET=your-secret \  ghcr.io/skyrxin/sast-mcp-server:full --transport streamable-http
# …or use the bundled compose file (exposes /health /ready /metrics too)docker compose up

Prefer to build locally?

docker build -t sast-mcp-server .                       # minimaldocker build -f Dockerfile.full -t sast-mcp-server:full .   # full

CodeQL and OWASP ZAP are not bundled in the image — CodeQL ships a multi-GB bundle (mount at runtime) and ZAP's run_active_scan orchestrates Docker on the host.

Reliability

"Production-ready" isn't a slogan here — it's measured in CI on every push:

• 226 tests, 74% line coverage (Codecov), green across Python 3.10–3.13, with mypy type-checking enforced.
• Self-scan, every build. The server runs its own scanners against its own code in CI; the SARIF + summary are published as the self-scan-report artifact. A snapshot lives in examples/self-scan/.
• Load-tested HTTP transport. scripts/loadtest.py runs in CI against the Streamable HTTP server: a local run sustains ~210 req/s at p95 ≈ 290 ms with zero failures across /health, /ready, /metrics under 30 concurrent workers.
• Ops endpoints — /health (liveness), /ready (cached scanner inventory, 503 when none installed), /metrics (Prometheus text).
• Bounded under load — a configurable concurrency cap (SAST_MCP_MAX_CONCURRENT_SCANS) around subprocess scanners, per-client token-bucket rate limiting (SAST_MCP_RATE_LIMIT_PER_MIN) for HTTP transports, per-scanner timeouts, and an optional incremental-scan cache.

Run the load test yourself:

python scripts/loadtest.py --requests 2000 --concurrency 50

Configuration

Environment Variables

Operational endpoints (HTTP transports)

When running with --transport streamable-http, the server exposes:

Incremental scans

scan_vulnerabilities and scan_all accept use_cache=true: the server fingerprints the target's files and reuses the previous scan when nothing has changed, so repeated scans in a session are fast.

Development

# Clone and install with dev dependenciesgit clone https://github.com/Skyrxin/sast-mcp-server.gitcd sast-mcp-serverpip install -e ".[dev]"
# Run testspytest tests/ -v
# Lintruff check sast_mcp_server/
# Run locallypython -m sast_mcp_server

Project Structure

sast_mcp_server/├── __init__.py          # Package version├── __main__.py          # python -m entry point├── server.py            # FastMCP server with all tools + /health /ready /metrics├── models.py            # Typed data models (Finding, Severity, etc.)├── config.py            # Central validated settings (env-driven)├── sarif.py             # SARIF 2.1.0 export and parsing├── aggregator.py        # Multi-scanner parallel execution + deduplication├── cache.py             # Scan caching, baselines, comparison, fingerprints├── auth.py              # JWT / API key authentication for remote transports├── metrics.py           # In-process Prometheus metrics├── ratelimit.py         # Per-client token-bucket rate limiting├── prompts.py           # MCP prompt templates (security workflows)├── resources.py         # MCP resources (sast:// dashboards and metadata)├── scanners/│   ├── base.py          # Abstract scanner base class│   ├── factory.py       # Scanner registry and factory│   ├── bandit.py        # Bandit (Python)│   ├── njsscan.py       # njsscan (JavaScript)│   ├── bearer.py        # Bearer (multi-language)│   ├── semgrep.py       # Semgrep (30+ languages)│   ├── trivy.py         # Trivy (CVEs, secrets, IaC)│   ├── codeql.py        # CodeQL (deep semantic SAST)│   ├── checkov.py       # Checkov (IaC policies)│   ├── gitleaks.py      # Gitleaks (git history secret scanning)│   ├── osv_scanner.py   # OSV-Scanner (SCA / dependency CVEs)│   ├── grype.py         # Grype (SCA + container image scanning)│   └── zap.py           # OWASP ZAP (DAST via Docker)├── enrichment/│   ├── ast_context.py   # AST-aware code context extraction│   ├── git_diff.py      # Git diff for incremental scanning│   ├── ignore_manager.py # Finding ignore list management│   ├── patch_prompt.py  # Builds LLM prompts to fix a cached finding│   └── patch_apply.py   # Applies/reverts agent-generated diffs via `git apply`├── reporting/│   ├── sbom.py          # CycloneDX SBOM/VDR (+ optional Syft inventory)│   ├── spdx.py          # SPDX 2.3 SBOM│   ├── vex.py           # CycloneDX VEX statements (triage decisions)│   ├── html.py          # Standalone HTML executive report│   ├── pdf.py           # PDF report (optional [pdf] extra)│   └── compliance.py    # OWASP / SANS / PCI / CIS mapping└── integrations/    ├── defectdojo.py    # Upload SARIF to DefectDojo    ├── github.py        # GitHub Code Scanning + PR comments    ├── gitlab.py        # GitLab MR comments    ├── slack.py / teams.py  # Webhook notifications    └── jira.py          # Create Jira issues

License

MIT