接入 AngusSecurity MCP：在 IDE 聊天中查詢高危發現、掃描結果及閘門

Once AngusSecurity MCP is connected, you can ask for scan conclusions right in IDE chat: any highs on this check, how to fix a finding, whether the gate will block, whether a report is ready—look up highs and alerts, explain findings, compare two checks, read or download report snapshots; with matching permissions and write on, you can also scan the current repo, image, or artifact. What comes back is a redacted conclusion, not a full raw scan dump into the model.

Main practical scenarios:

• Before a release: any Critical / high findings on this check?
• Point at one finding: what does the rule mean, and how do I fix it?
• Diff two scans: what’s new, what’s gone, what’s still there?
• Simulate with an existing policy: will the gate block this result?
• Read an existing report or download its body; with enough permissions, create a report for a finished check
• When allowed, scan the current repo (or an image / artifact)

Takeaway: Security MCP exposes up to sixteen tools—most query findings, alerts, CVEs, licenses, gates, and reports; with permissions and write on, you can also scan or create a report for a finished check. No suppress of false positives, no secret plaintext, no publish/share. Use a portal-issued user access token for this product.

Console and host wiring: Connect AngusSecurity with MCP in the references. “Which repo holds this coordinate / client setup” is AngusRepo MCP—eight tools, write always off. Here the answer is scan conclusions, CVEs, licenses, gates, and report snapshots. Unified engines and how reports export: replace scanner scripts; how to export scan reports.

1. Six questions agents keep asking (and what is out of scope)

1. Scan the current repository / image / artifact / named files.
2. Read that check’s issues grouped by code security, dependencies, secrets, and licenses.
3. List or explain a finding (secret snippets stay hidden).
4. Look up a CVE in the local vuln DB; summarize visible licenses.
5. Simulate a gate with an existing policy; compare two completed checks.
6. Read an existing report or download its body; with enough permissions, create a report for a finished check.

Explicitly out of scope: suppress false positives, reveal secret plaintext, change alert status, edit policies, publish/share/delete reports, manage credentials and engines, upload a whole unnamed tree, export engine raw output.

2. Shape: in-process endpoint + loopback

Chat only sees named tools; the server loops back in-process to the existing management API so permissions and audit stay one path.

IDE / Agent
  --POST {gateway}/security/mcp-->  Boot /mcp
  --loopback Bearer-->  /api/v1/** (named semantic tools)

A separate Node that re-wraps REST would duplicate policy. Loopback has no generic path parameter, so the model cannot invent the general upload gateway, reveal, or publish routes.

Rate limit: at most 30 successful calls per account per minute; then rate_limited—wait briefly.

3. Sixteen tools (names as in the docs)

With write on and full related permissions, an analyst or admin should see exactly 16 on tools/list. Viewers see fewer—§4.

3.1 Start a scan and read that result

3.2 Day-to-day lookup

Pagination starts at 1; at most 20 per page.

3.3 Compare, vuln DB, and gates

3.4 Reports

Explicitly not in the list: suppress, reveal, publish, share, delete reports, edit policies, upload a whole unnamed tree, export engine raw output.

A report is a snapshot at generation time and does not rewrite when the next scan finishes—same stance as “how to export scan reports” in the references.

4. Which tools each role sees

If the account lacks permission, the tool does not appear; calling it by name yields permission_denied. MCP does not grant extra access. Try the same action in the console with the same account before blaming MCP.

5. Auth boundaries (easy to get wrong)

Visibility matches what that user sees on /api/v1, same as the console. Personal tokens: avatar → Me (/profile). Use a service account for CI; do not commit a personal session into a repository.

6. Host wiring (minimal copy-paste)

Keep the token in an environment variable; do not commit it:

{
  "mcpServers": {
    "angus-security": {
      "url": "https://<gateway-host>/security/mcp",
      "headers": { "Authorization": "Bearer ${ANGUS_SECURITY_TOKEN}" }
    }
  }
}

Public shape is gateway {host}/security/mcp. A direct private process uses https://<security-host>/mcp (default port 1811; behind a reverse proxy you need not expose that port publicly). Copy the URL from System → MCP in the console.

Stdio-only hosts forward with @anguskit/mcp --product security (reads ANGUS_TOKEN / ANGUS_BASE_URL, not the same names as ANGUS_SECURITY_TOKEN):

{
  "mcpServers": {
    "angus-security": {
      "command": "npx",
      "args": ["-y", "@anguskit/mcp", "--product", "security"],
      "env": {
        "ANGUS_BASE_URL": "https://<gateway-host>",
        "ANGUS_TOKEN": "${ANGUS_SECURITY_TOKEN}"
      }
    }
  }
}

ANGUS_BASE_URL is the gateway origin without /security/mcp. Without a product prefix in private deploy, use --endpoint + https://<security-host>/mcp. If the default npm registry lacks the package, the docs suggest https://repo.anguskit.com/npm/npm-public/. Full host table: MCP docs in the references.

6.1 Dialogue examples: question → tool → expected fields

Example A: scan, then ask for issues

Example B: explain an existing finding

Example C: gate and report

After connect, also tools/list against the role table, then smoke-test security_explain_finding on a finding that already exists in the console.

7. Short industry shape comparison (and vs Repo)

Comparison explains product trade-offs, not a buying rank. Details follow each vendor’s docs of the day.

8. Boundaries and failure table

Three common failures after connect: the wrong product’s token won’t get in, an empty finding list rewritten as “safe,” and asking for report body while the scan or report is still running—see the table and FAQ below.

FAQ

Will MCP reveal secrets or mark false positives?

No. Explain and alerts keep masks; suppress, reveal, and approval stay in the console. One reason: some suppress scopes auto-approve—if the model frames “accept risk” as a fix, it would write an effective suppress.

Can I still read / download reports when write is off?

Queries need Reports:VIEW. Downloads need Reports:EXECUTE and do not use the write switch. What the write switch removes is security_scan and security_create_report.

How does this relate to CLI / CI?

CLI and pipelines remain the main path for batch runs and gates; MCP brings the same jobs and findings into IDE chat. Unified engines: replace scanner scripts in the references.

Why not a generic read-only gateway?

Named tools keep the loopback whitelist small, avoid exposing the whole admin API to the model, and make projection redaction easier.

Why can’t other-product tokens / repod_ connect?

MCP uses this portal’s user access token with the same rights as the console. Treating automation or cross-product tokens as chat identity widens the surface—401 before auth proceeds.

Why can’t security_license_summary answer “this job”?

It rolls up licenses visible to the account and cannot bind a jobId. For one job’s license issues, use security_issues.

Can an empty list be called “safe”?

No. Also distinguish: job unfinished, a capability with notScanned=true, and filters that missed. None of those mean the target is safe.

Takeaways

• Differentiated from Repo MCP: here it is scan conclusions / gates / reports; write surface is only scan + create report, gated by switch and role.
• Wiring: gateway /security/mcp + PAT; repod_ and other-product tokens will not work; secrets masked; empty ≠ safe.
• For wiring details, follow the docs and System → MCP in the console.

Reference articles

• Connect AngusSecurity with MCP: https://www.anguskit.com/en/docs/security/get-started/mcp
• Replace Semgrep, Trivy, and Gitleaks scripts: https://www.anguskit.com/en/blog/replace-semgrep-trivy-gitleaks-with-one-scan
• How to export scan reports: https://www.anguskit.com/en/blog/security-scan-report-snapshot-export
• Capability-scoped reports: https://www.anguskit.com/en/blog/capability-security-reports-sast-sca-secret-license
• Intranet vuln DB update and rollback: https://www.anguskit.com/en/blog/intranet-vuln-db-offline-update-rollback
• Audit evidence checklist: https://www.anguskit.com/en/blog/compliance-audit-security-scan-evidence
• Read-only artifact queries (AngusRepo MCP): https://www.anguskit.com/en/blog/angusrepo-mcp-server-design