Deep SAST MCP Server — hosting & platform registration

A Model Context Protocol server that gives the agent github_code_security_assessment_Piyush_tiwari deterministic, 100%-file-coverage security findings by wrapping Semgrep (SAST), gitleaks (secrets) and osv-scanner (dependency CVEs).

Why: the agent's GitHub tools sample code; this server PARSES every file. The LLM then triages real findings instead of grepping — that's the thoroughness unlock.

Files

• server.py — the MCP server (FastMCP). 7 tools: scan_repository, get_scan_status, list_findings, get_finding_context, get_file, get_dependency_report, cleanup_scan.

• requirements.txt — Python deps (fastmcp, semgrep).

• Dockerfile — image with all three scanners on PATH.

Related MCP server: Security Scanner MCP Server

Tools exposed to the agent

Normalized finding fields: id, scanner, rule_id, title, severity, owasp, cwe, path, start_line, end_line, snippet, fix_hint. (gitleaks snippet is redacted.)

1. Run locally (smoke test)

cd mcp-server
pip install -r requirements.txt
# also install the binaries locally if not using Docker:
#   brew install gitleaks osv-scanner   (mac)  — or download release binaries
python server.py --transport http --host 127.0.0.1 --port 8080

The server prints its MCP HTTP endpoint (e.g. http://127.0.0.1:8080/mcp).

2. Build & run with Docker (recommended)

cd mcp-server
docker build -t deep-sast-mcp .
docker run --rm -p 8080:8080 \
  -e MAX_REPO_MB=500 -e SCAN_TIMEOUT_S=1800 \
  --read-only --tmpfs /tmp \
  --security-opt no-new-privileges \
  deep-sast-mcp

Notes: --read-only + --tmpfs /tmp keep clones in memory-backed tmp and block writes elsewhere. The container runs as non-root.

3. Host with a public HTTPS endpoint (required by the platform)

The platform can only register tools at a public https:// URL. Pick one:

• IBM Code Engine: ibmcloud ce application create --name deep-sast-mcp \ --image <registry>/deep-sast-mcp --port 8080 --min-scale 0. It gives an https URL.

• OpenShift / ROKS: deploy the image, expose a Route (TLS edge), use the Route host.

• Any container host behind an HTTPS load balancer.

Harden the endpoint:

• Put it behind an auth token (bearer) or mTLS. Add a check in server.py if needed.

• Restrict egress: the server should only reach github.com/gitlab.com/bitbucket.org (clone) — block everything else to limit SSRF/exfil from a malicious repo.

• Set CPU/mem limits and keep SCAN_TIMEOUT_S / MAX_REPO_MB sane.

4. Register the MCP server in IBM Consulting Advantage

The platform runs IBM ContextForge MCP Gateway (github.com/IBM/mcp-context-forge). You register our server there as a new MCP gateway.

Path: open your Agentic App → Tools tab → "Access MCP Gateway". This opens the ContextForge admin ("Gateway Administration"). Go to MCP Servers (the #gateways section) → "Add New MCP Server or Gateway" and fill:

After adding, ContextForge federates the server, its 7 tools appear under Tools, and you can group them into a Virtual Server (with its own API key) that the app's agents consume.

Note: existing team servers show URLs like https://servicesessentials.ibm.com/mcp-gateway/service/gateway/servers/<id>/mcp — that is the gateway's federated proxy URL it assigns AFTER you register your real backend URL. Your backend (this server) must be reachable over public HTTPS for the gateway to reach it.

5. Wire it into the agent / multi-agent app

• Single agent: Edit github_code_security_assessment_Piyush_tiwari → Add tools → select the Deep SAST tools → Republish. Update instructions to prefer scan_repository for coverage, then list_findings + get_finding_context to triage.

• Multi-agent app (Agentic App Studio): give scan_repository/list_findings/ get_finding_context/get_file/get_dependency_report to the Inventory + Reviewer agents (see ../design/multi-agent-app-spec.md). The Inventory agent calls scan_repository; reviewers consume list_findings filtered by severity/path.

6. Agent workflow once registered

1. scan_repository(url) → wait for state=done (report files_scanned/total).

2. list_findings (paginate; filter by severity) → the complete finding set.

3. For each finding → get_finding_context → confirm source→sink, drop false positives, finalize severity + remediation.

4. get_dependency_report → supply-chain findings; confirm notable CVEs via Web Search.

5. Emit the report + a COVERAGE LEDGER backed by REAL scanner numbers (not an estimate).

6. cleanup_scan(scan_id) when done.

Security notes (do not weaken)

• Scanners parse, never execute, the target code.

• Per-scan temp workspace, deleted by cleanup_scan.

• Clone is shallow, host-allowlisted, and size-capped (MAX_REPO_MB).

• gitleaks secret VALUES are redacted before leaving the process.

• get_file blocks path traversal outside the scan workspace.