security-recipes.ai

Search CVEs. Remediate vulnerabilities with AI agents. Sourced facts stay sourced, remediation stays bounded, and every plan carries verification, rollback, and stop conditions — the live site's contract, and this repo's.

security-recipes.ai is an Eleventy site for sourced CVE intelligence and evidence-gated vulnerability remediation that AI agents can consume without inheriting deployment or production authority.

The project is intentionally narrow:

• a complete rolling Medium/High/Critical CVE database,
• evidence-qualified canonical CVE remediation records,
• practical security remediation recipes,
• prompt and rules-file examples,
• agent setup guides,
• MCP integration patterns,
• an optional read-only MCP server for recipe search and approved upstream MCP context,
• a reusable GitHub Action that turns this guidance into toggleable CI health checks.

It is not a scanner, ticketing system, SOAR platform, deployment tool, or custom security toolkit. Existing security tools should produce the findings; this site helps agents use the right remediation context and stop at the right time.

Start with the live CVE Database for an exact vulnerability or the AI Vulnerability Remediation Playbooks for the evidence-to-patch workflow. Agent-specific guides cover Codex, Claude Code, Cursor, GitHub Copilot, Devin, Shiba Studio, Hermes Desktop, and OpenClaw. The Visual Guide shows the complete path from source qualification and search discovery to a bounded plan, proof, rollback, and human review. For the distinct problem of securing an agent system's identities, tools, connectors, context, memory, runtime, and recovery controls, use AI Agent Security.

Current product and workflow

Qualified search discovery

The complete catalog remains searchable, while public canonical CVE pages stay limited to reviewed or evidence-qualified records. Those pages ship unique search metadata, server-rendered core facts and affected-version evidence, one remediation authority (stable reviewed guidance first, otherwise complete source-linked AI enrichment), a short approval-gated AI implementation prompt, canonical URLs, breadcrumbs, and Article/TechArticle structured data. The CVE database describes the catalog as a Dataset; the remediation pillar exposes its visible seven-step workflow as a HowTo. Year-partitioned CVE sitemaps contain only indexable canonical routes, and the build fails when sitemap parity, canonical ownership, crawl reachability, metadata limits, or same-origin links drift.

Indexability is also withheld from mass-templated recipe children. The 72 development code-hygiene recipes and 39 generated compliance-framework recipes remain browsable from their canonical hubs with noindex,follow while they share a common method. A bounded rendered-body similarity gate prevents a child from re-entering sitemaps until its evidence, examples, and tests are materially distinct. The hubs remain indexable and carry the shared discovery context.

After an SEO-bearing release, the public revision must match the merge commit before sitemap submission or URL inspection. The Caddy deployment guide documents the DNS-verified Search Console handoff, priority live-URL checks, sitemap submission, indexing requests, and query monitoring. Submission is a discovery hint; it does not guarantee indexing or a particular ranking.

The remediation pillar also records a public repository example for CVE-2026-13149 in brace-expansion. It ties the dependency-only change to the reviewed pull request, tests, advisory evidence, and recovery path while explicitly separating the same PR's unrelated Fail2Ban work.

CVE search to canonical record	CVE evidence to bounded agent plan
Proof and human review	Read-only MCP context

What this project is for

AI coding agents can help close security findings when their work is bounded: one finding, one recipe, one reviewed output.

security-recipes.ai helps teams answer:

• Which recipe matches this finding?
• Which prompt should the agent use?
• Where do I put the instructions for Copilot, Claude, Cursor, Codex, or Devin?
• Which MCP servers should the agent read for advisory, scanner, repository, or runbook context?
• What should the PR or triage note include before a reviewer trusts it?

What ships

• Eleventy documentation site (fast static builds, no Go toolchain).
• CVE-first observatory home page and data-first CVE database.
• Recipe hubs for dependency, SAST, sensitive-data, base-image, CVE, and default-hardening remediation.
• CVE intelligence intake policy, prompt, fixtures, and evaluator for routing advisory signals before an agent patches.
• A complete rolling ten-year Medium/High/Critical CVE catalog composed from integrity-verified NVD JSON 2.0 feeds, CISA KEV metadata, and every applicable vetted remediation archetype. Only reviewed stable Markdown pages override that conservative baseline.
• An integrity-hashed search allowlist that publishes canonical CVE pages only for reviewed stable Markdown or AI enrichment that passes the deterministic recipe-ready evidence contract. The full database remains searchable even when a record is not eligible for search indexing.
• A versioned seven-phase agentic change contract for every catalog CVE: discover, assess, mitigate, remediate, verify, rollback, and triage. Each action declares likely file targets, mutation and approval boundaries, required evidence, outputs, and failure behavior without guessing a patch or fixed version.
• A structured compliance library spanning 39 security, privacy, assurance, resilience, and software-supply-chain frameworks without reproducing licensed control text. Its framework hub is the search surface; templated child assessments remain noindex,follow until differentiated.
• A 72-recipe code-hygiene library covering cross-language and ecosystem- specific audit, remediation, verification, and stop-condition workflows. Its development children remain noindex,follow while their bodies share a generated template.
• Recipes with existing prompt collections preserved.
• Agent setup guides for GitHub Copilot, Claude, Cursor, Codex, and Devin.
• MCP integration guidance for public and organization-approved security data sources.
• Optional read-only FastMCP server in mcp_server.py for recipe search, retrieval, and opt-in upstream MCP context.
• Docker and Docker Compose configuration for local or droplet hosting.
• Helper scripts for site maintenance, validation, imports, and deployment.

Repository map

Path	Purpose
content/	Recipes, documentation, remediation guides, and agent setup pages.
eleventy.config.js	Site build configuration (permalinks, feeds, tag pages).
_includes/	Page layouts: docs chrome and the standalone home page.
lib/	Build modules: shortcode ports, JSON feed builders, SEO head.
assets/	Site CSS and JavaScript for the recipe browser, navigation, and helper tools.
static/	Images, logos, schemas, and static assets.
static/api/cve-catalog/	Complete sharded CVE catalog, year-partitioned machine index, compressed browser-search index, provenance manifest, and archetypes.
data/cve/	Human-reviewed remediation archetypes, deterministic AI-enrichment cache, and generated-recipe ownership ledger.
data/compliance-frameworks/	Structured compliance-framework catalog and source registry.
data/code-hygiene/	Structured code-hygiene catalog, source registry, and routing fixtures.
docs/	Repository documentation and legacy screenshot assets; current README and visual-guide images live in static/images/.
mcp_server.py	Optional read-only MCP server for recipe search and approved upstream MCP context.
mcp-server.toml.example	MCP server configuration template.
Dockerfile	Site image.
Dockerfile.mcp-server	Optional MCP server image.
docker-compose.yml	Production-style local stack.
scripts/	Helper scripts for maintenance and deployment.

Core content areas

• CVE Database: sourced CVE intelligence, affected-version evidence, and canonical remediation records.
• AI Vulnerability Remediation: evidence-gated playbooks from one finding to a reviewed patch or triage note.
• AI Agent Security: threat modeling, production baselines, source boundaries, control routing, evidence, and incident readiness for the AI-agent system itself.
• Quick Start: one finding to one reviewed PR or triage note.
• AI Agent Comparison: verified operating modes, native instructions, expected artifacts, prerequisites, and review gates for Copilot, Claude Code, Cursor, Codex, and Devin.
• Recipes: reusable prompts, instructions, rules, skills, and review checklists.
• MCP Integration: how to connect security context safely.
• Visual Guide: the qualified search-discovery, CVE-to-plan, proof, rollback, review, and read-only MCP flow in five diagrams.
• Docs: site usage, agent consumption patterns, and contribution guidance.

Python remediation tooling

The Python suite is an optional execution companion to the documentation. It can inspect a bounded workspace, select any of the 75 remediation playbooks, create a durable run packet, record integrity-hashed evidence, and verify the packet before agent or reviewer handoff. It remains local and conservative: it does not merge code, deploy changes, or call external systems on its own.

python scripts/security_recipes_remediation_suite.py playbook list
python scripts/security_recipes_remediation_suite.py playbook inspect \
  --playbook vulnerable-dependencies --workspace .
python scripts/security_recipes_remediation_suite.py playbook start \
  --playbook vulnerable-dependencies --workspace . \
  --finding finding.json --run-dir .security-recipes/runs/dependency-fix
python scripts/security_recipes_remediation_suite.py playbook verify \
  --run-dir .security-recipes/runs/dependency-fix

The repository also includes domain-specific generators and evaluators for playbooks that need richer evidence packs or runtime policy decisions. The site and JSON registry remain useful without Python; the tools make the same workflow contracts directly executable by CI, orchestrators, and approved coding agents.

Deployment helpers worth knowing:

• scripts/setup_digitalocean_droplet.sh: Ubuntu droplet bootstrap with Docker, host hardening, and optional Caddy-managed HTTPS.
• scripts/configure_nginx_letsencrypt.sh: host nginx reverse proxy setup for teams that want Let's Encrypt on nginx instead of Caddy.
• README.nginx-letsencrypt.md: operator-focused walkthrough for the nginx deployment path.

Recommended operating model:

1. Let existing SCA, SAST, secrets, CI, cloud, and ticketing systems produce findings.
2. Attach a matching security-recipes.ai recipe and prompt.
3. Let the agent read only the files and MCP context needed for the finding.
4. Require tests and human review before merge.
5. Keep broad automation, write access, and deployment outside the first loop.

Guidebook and execution tools

The site is a guidebook for remediation work: recipes, prompts, agent setup, MCP/API integration notes, and review patterns. Runtime automation belongs in the user's approved agent host, CI system, ticketing workflow, or scanner platform rather than a site-hosted chatbot.

Python tools in scripts/, tools/, and mcp_server.py support maintainers and self-hosters with playbook execution packets, evidence verification, domain-specific evaluation and generation, validation, advisory import, recipe search, and optional read-only MCP access.

Optional MCP server

The MCP server is read-only by default. Its baseline role is to let MCP-compatible agents search and retrieve recipes. Self-hosted deployments can also configure it as a context hub for approved upstream MCP servers without putting those credentials into the public site.

Retrieved context never grants mutation authority. Any connector that can change repositories, tickets, secrets, deployments, or production systems must be configured and approved separately by the calling host.

Common tools:

• recipes_search
• recipes_list
• recipes_get
• recipes_cve_catalog_info
• recipes_cve_search
• recipes_cve_get
• recipes_match_finding
• recipes_playbooks_list
• recipes_playbook_get
• recipes_playbook_plan
• recipes_mcp_upstream_servers
• recipes_mcp_upstream_tools
• recipes_mcp_upstream_call
• recipes_mcp_upstream_context

The MCP server accepts both generated recipe feeds:

• /api/recipes.json is the preferred agent feed with category, severity, CVE/GHSA, ecosystem, and handoff metadata.
• /recipes-index.json remains supported for legacy consumers.
• /recipes-browser.json is the compact interactive-library feed. The /recipes/ page server-renders 18 crawlable recipe cards and an exactly matching hydration seed, then requests the complete feed only when a visitor focuses search, filters, sorts, follows a filtered URL, or loads more.

The complete CVE catalog is also available without MCP:

• /api/cve-catalog/manifest.json declares the exact date/severity policy, source hashes, coverage counts, and shard inventory.
• /api/cve-catalog/runtime-summary.json is the small browser bootstrap with coverage totals and content-derived cache versions for every runtime asset.
• /api/cve-catalog/index.json is a small manifest for the complete publication-year partitions under /api/cve-catalog/indexes/. Offline consumers can fetch only the years they need; neither a browser page load nor an exact MCP lookup parses those partitions.
• /api/cve-catalog/search is the bounded, same-origin broad-search endpoint. It is pinned to the shard-set revision declared by runtime-summary.json, rate-limited at nginx, and returns at most 100 previews. The production MCP image serves it from a read-only SQLite FTS database built and whole-file verified against the same manifest. Focus alone and an incomplete CVE-YYYY-NNNN identifier make no search request.
• /api/cve-catalog/records/{cve} is the bounded, same-origin exact-record endpoint. Every request pins the shard-set revision, and the MCP service verifies and opens only the one deterministic shard containing that CVE. Current browsers use this endpoint instead of learning the shard namespace.
• /api/cve-catalog/browser-index.json.gz remains for one compatibility window when an older runtime summary does not declare the search and record APIs. Current browsers do not download it when the APIs are declared, so visitors no longer pay the complete-corpus transfer or memory cost.
• Canonical CVE pages server-render their overview, affected-version evidence, selected remediation authority, AI implementation and verification handoff, sources, provenance, citation, and schema. They do not embed or hydrate the catalog application. A compact link to the exact gzip JSON Lines shard remains available for machine-readable provenance without adding a browser fetch.
• /api/cve-catalog/search-indexable.json is the compact, integrity-hashed allowlist for canonical CVE pages, related-CVE links, and search discovery. Its policy accepts only reviewed stable Markdown or complete AI enrichment that passes the deterministic recipe-ready evidence contract. Every browser result links to its local /cve/<ID>/ record. Allowlisted records are materialized as indexable static pages; all other records use the bounded runtime renderer with noindex,follow and retain their official CVE.org source in the record.
• /api/cve-catalog/archetypes.json contains the reviewed remediation contracts used to compose a conservative recipe for every catalog record. It also contains the versioned agentic action schema and ecosystem-specific file-target hints shared by the browser and MCP server.

Canonical CVE pages use one primary-reference set for the visible source list and structured-data citations. Raw generated records admit NVD, CVE.org, scoped CISA KEV records, and source-linked vendor advisories, patches, release notes, or mitigations; broken, third-party-only, exploit-only, and generic vulnerability-database links are not promoted automatically. Stable reviewed Markdown can deliberately cite additional HTTPS evidence in its References section. When remediation spans several supported branches or product families, the displayed action preserves every trusted fixed-release claim instead of collapsing the guidance to one incomplete upgrade.

Development and catalog-owned stable CVE Markdown emit no standalone page in the pure static build and are excluded from Eleventy and generic recipe/search feeds, tag pages, RSS, and the sitemap. The three pre-catalog historical stable recipes remain ordinary rendered content. Production can retain a legacy recipe URL as a redirect to the canonical CVE route through nginx and the MCP-backed landing service. Use the dedicated catalog or recipes_cve_* MCP tools for complete discovery.

The browser's exact-ID path and revision-pinned search API cover every in-scope Medium, High, and Critical record declared by the manifest. The MCP server exposes the same SQLite-backed coverage through recipes_cve_search; a successful recipes_cve_get returns the normalized source record, source identifiers and references, applicable archetypes, composed remediation contract, and a self-contained agentic_change_plan. The plan expands each mitigation and remediation instruction into ordered code/file operations with verification, rollback, evidence, approval, and triage requirements. It also preserves explicit CPE truncation metadata when the source match set exceeds the bounded record.

Daily CVE synchronization and optional AI enrichment

.github/workflows/cve-catalog-sync.yml runs every day at 09:23 UTC and can also be dispatched manually. It verifies and joins the NVD JSON 2.0 annual feeds and CISA KEV catalog, regenerates every catalog index/shard, validates the result, refreshes recipe-derived deterministic evidence in dependency order, runs the catalog tests, and opens or refreshes automation/cve-catalog-sync as a pull request to the default branch. Repository Settings > Actions > General > Workflow permissions must allow GitHub Actions to create pull requests for first-run PR publication.

Set CVE_AUTO_MERGE_ENABLED=true to deliver a safety-approved catalog PR after its exact head revision passes the dedicated validation workflow. The merge job reads CVE_SYNC_APP_ID and CVE_SYNC_APP_KEY from the cve-sync environment, mints a scoped GitHub App token, and recommits the catalog tree so ruleset 18147261 sees an attributed stevologic-cve-sync commit. That ruleset currently has no bypass actors and treats GITHUB_TOKEN / github-actions commits as unattributed; the App must be added as a PR-only bypass actor before merge can succeed. Missing App credentials fail the merge job on the first run. After a guarded App merge it verifies that the returned merge SHA is still current main, then the App push emits the real Build. The production deploy gate recognizes only those CVE-qualified Build dispatches, so scheduled monitors and unrelated manual workflows cannot deadlock or satisfy a release.

The source sync does not require a secret. Catalog enrichment defaults to xAI grok-4.7 through the existing XAI_API_KEY Actions secret. One repo variable, CVE_ENRICH_MODEL, selects the provider and model. Leftover-gold review, content refresh, AI maintenance, AI issue maintenance, and this repository's security-health action still use Grok through XAI_API_KEY.

gh secret set XAI_API_KEY --repo stevologic/security-recipes.ai
gh secret set OPENAI_API_KEY --repo stevologic/security-recipes.ai

The catalog workflow defaults to grok-4.7 and at most 20 new or source-changed records per run. The scheduled queue is derived from the tracked NVD/CISA catalog: a candidate must have a valid tagged vendor advisory, patch, release-note, or mitigation URL. Source-complete records remain eligible because they still need a sourced remediation synthesis; within each KEV and severity band they rank ahead of records with deterministic source gaps, followed by affected-product/version evidence and recency. This uses the existing daily request budget and does not require an additional manual run. The model and limit can be changed with optional Actions variables; the enrichment limit is hard-bounded from 0 to 50:

gh variable set CVE_ENRICH_MODEL --body "grok-4.7" --repo stevologic/security-recipes.ai
gh variable set CVE_ENRICHMENT_LIMIT --body "20" --repo stevologic/security-recipes.ai

CVE_ENRICH_MODEL is the only switch. grok-* uses xAI and XAI_API_KEY. gpt-* uses OpenAI and OPENAI_API_KEY. To move to Daybreak later, change that one variable to gpt-6-sol+daybreak_blue. A workflow_dispatch input daybreak_probe=true makes one probe call for whichever provider is configured and skips the catalog sync.

AI output is supplemental and explicitly labeled. It uses strict structured output, only cites URLs actually returned in the Responses API web-search provenance, and is stored reproducibly in data/cve/ai-enrichments.json. A complete enrichment becomes a CVE-specific Markdown draft only when a separate gate finds claim-level affected-product, exposure, remediation, and verification evidence tied to the exact URL of a tagged trusted advisory reference. Every required claim must independently meet that rule, and every generated recipe requires a cited, concrete fixed-version claim.

Cached enrichment is re-evaluated instead of becoming permanent: recipe-ready entries become refresh candidates after 30 days, KEV entries after 60 days, and other complete/not-specific or insufficient-evidence entries after 180 days. A manually prioritized CVE forces a refresh inside the existing request cap. The last valid cached result stays attached if that refresh fails; an invalid source fingerprint remains fail-closed. The synchronization report and automation-health summary expose refresh-due and manually prioritized counts.

Eligible drafts are written as maturity: development files named content/recipes/cve/ai-enrichment-cve-*.md. They stay outside generic recipe discovery and never override a stable reviewed recipe. A human reviewer can set ai_enrichment_review_status: human-reviewed-development-draft to withhold an otherwise evidence-ready enrichment from public remediation authority, or ai_enrichment_review_status: approved-for-ai-authority to approve that use. Unannotated generator-owned drafts retain the automated evidence gate, while stable Markdown always wins. The ownership ledger in data/cve/ai-generated-recipes.json records each generated file hash; automation may refresh or remove only an untouched hash-matching draft. A human edit, or any existing human development/stable recipe for the same CVE, makes that Markdown human-owned and blocks automated replacement. AI generation never changes source CVSS/KEV facts, affected-version data, archetype selection, or reviewed stable Markdown. A missing key, API refusal, timeout, or rate limit does not block the NVD/CISA refresh; calls stop after three consecutive failures or a 15-minute budget, and valid cached enrichments remain attached. A manual run may prioritize named CVEs, but those IDs consume slots inside that run's existing cap and never bypass the recipe-ready evidence gate:

gh workflow run cve-catalog-sync.yml --ref main \
  -f ai_enrichment_limit=20 \
  -f priority_cve_ids="CVE-2026-58644,CVE-2026-56164"

A manual dispatch is an additional workflow run and can therefore make additional requests; it is not needed for the daily deterministic queue. A manual run on a non-default branch uploads its enrichment cache, ownership ledger, and generated drafts as a short-lived workflow artifact for review.

.github/workflows/leftover-review.yml runs every day at 13:17 UTC and live-verifies leftover-gold CVE leftovers against GitHub Advisories and NVD. Leftover-gold criticals and highs drain first. After those close, each run reviews up to 25 leftover-gold medium and low pages, records completed IDs in data/cve/leftover-review-state.json, and opens a labeled auto-merge PR. The leftover-review job uses the Grok Build CLI with XAI_API_KEY and no-ops when that secret is missing or the leftover-gold queue is empty.

The runtime paths are deliberately bounded for catalog-scale traffic:

• the hub bootstraps from the compact runtime summary, exact lookups call the revision-pinned same-origin record API, and title/product/vendor/filter search calls the search API only after explicit search intent;
• broad search returns at most 100 previews from immutable read-only SQLite, has a three-second HTTP boundary, and never decodes the complete catalog in a visitor process or on the browser main thread;
• the exact-record service verifies and opens one shard per request; MCP exact retrieval uses the same shard-only path, while non-exact text search uses the manifest-pinned SQLite database behind a dedicated executor, bounded admission queue, query deadlines, and nginx rate limit;
• immutable browser cache keys come from the declared record/search contract, archetype hash, and shard-set revision rather than an upstream timestamp.

The implemented build boundary, exact-shard delivery model, evidence-gated SEO policy, SQLite search runtime, and remaining artifact-publication migration are documented in CVE scale architecture.

The production image builds the SQLite artifact once in its cached image layer, records its independent SHA-256 sidecar, and validates schema, catalog revision, record count, manifest digest, file digest, and representative FTS postings at startup. RECIPES_MCP_EAGER_CVE_SEARCH now applies only to the legacy local fallback when no SQLite path is configured. For sustained search traffic, run multiple paired MCP instances; exact shard reads remain isolated from the bounded text-search executor and queue.

Run npm run icons after changing the site mark. It regenerates the opaque Apple touch icon and the 192/512/maskable installed-app assets checked by the production performance gate.

Production builds precompress large JSON/XML feeds for nginx gzip_static, validate stable/draft discovery boundaries, and enforce payload/file-count budgets with npm run check:performance.

Run it with Docker:

docker build -f Dockerfile.mcp-server -t security-recipes-mcp .
docker run --rm -p 8123:80 security-recipes-mcp

Connect an MCP client to:

http://localhost:8123/mcp

Run it locally with Python:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements-mcp-server.txt
python mcp_server.py

Windows PowerShell activation:

.\.venv\Scripts\Activate.ps1
python mcp_server.py

Run the site locally

Prerequisites:

• Node.js >= 20
• Python >= 3.10 with requirements-mcp-server.txt installed for the production npm run build CVE prerender step
• Git

python -m pip install -r requirements-mcp-server.txt
npm install
npm run serve

Open:

http://localhost:8080

npm run serve watches for changes and rebuilds incrementally. A one-off production build is npm run build (output lands in public/). The build performs a Python/dependency preflight before deleting an existing output and then uses the same CVE renderer as the MCP runtime. Eleventy deliberately does not passthrough-copy static/api/cve-catalog/: after page materialization, a bounded post-build step rejects links, orphan files, unsafe paths, and manifest byte/hash mismatches before installing that catalog subtree. Static assets outside the catalog, including root dotfiles, retain normal passthrough behavior.

For an isolated catalog build, set SECURITY_RECIPES_CVE_CATALOG_ROOT to its absolute publication directory. Eleventy data, qualified-page materialization, and the validated catalog copy all use that same root. npm run serve does not rerun the materializer or catalog copy, so run npm run build once first when you need canonical /cve/<ID>/ pages and the catalog API tree in the development server; later incremental rebuilds retain those post-build outputs.

Docker Compose

Create an environment file:

cp .env.example .env

Start the stack:

docker compose up -d --build

Use the Docker Compose v2 plugin (docker compose). The legacy Python docker-compose v1 package is not supported for this stack; it can crash with KeyError: 'id' while following logs or KeyError: 'ContainerConfig' while recreating containers on newer Docker Engine releases.

On Ubuntu/Debian hosts, install Compose v2 and a compatibility shim with:

sudo bash scripts/install_docker_compose_v2.sh

Default routes:

site: http://127.0.0.1:8080/
agent recipe feed: /api/recipes.json
MCP endpoint: /mcp

The Compose stack keeps the public site and its dynamic CVE/MCP renderer in matching blue/green pairs:

• security-recipes / mcp-server-blue: blue site and renderer.
• security-recipes-green / mcp-server-green: green site and renderer.
• mcp-server: transitional singleton retained for the first paired rollout and backwards-compatible manual Compose workflows. It reads the locally built site feed at http://security-recipes/api/recipes.json, so a fork or droplet serves its own recipes instead of depending on the public production index.

deploy.sh starts and revision-verifies the withdrawn slot's MCP container before its site container, validates a canonical CVE directly, and only then admits the pair to Caddy. Manual Compose startup retains the singleton default so the first rollout remains compatible with the previously installed script.

For an nginx or Caddy reverse proxy with Let's Encrypt, keep Docker bound to loopback and let the proxy own public ports 80 and 443:

SECURITY_RECIPES_HTTP_PORT=127.0.0.1:8080

Then proxy to:

location / {
    proxy_pass http://127.0.0.1:8080;
    proxy_set_header Host $host;
    proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
    proxy_set_header X-Forwarded-Proto $scheme;
}

If you want a turnkey host nginx + Let's Encrypt setup, run:

sudo bash scripts/configure_nginx_letsencrypt.sh \
  --domain security-recipes.ai \
  --email [email protected]

The full operator guide lives in README.nginx-letsencrypt.md.

DigitalOcean droplet

For a fresh Ubuntu droplet, use the helper script:

sudo bash scripts/setup_digitalocean_droplet.sh \
  --domain security-recipes.ai \
  --email [email protected]

The script installs Docker/Compose, configures a locked app user, enables basic host hardening, starts the Compose stack, and can place Caddy in front for HTTPS. It also enables a Caddy-aware Fail2Ban jail: five final HTTP 404 responses for high-confidence exploit-probe paths (for example .env, Git, WordPress, phpMyAdmin, or PHPUnit probes) from one client within five seconds block that address from the site's TCP and HTTP/3 ports for one hour, after which access is restored automatically. Ordinary missing pages, CVE-shaped misses, and archive pagination misses do not consume the ban budget.

Point both the apex and www DNS records at the Droplet before setup. Managed Caddy obtains certificates for both names and permanently redirects www to the apex canonical host; redirecting only at HTTP would leave HTTPS crawlers unable to complete the TLS handshake. The same stack can serve security-recipes.si and www.security-recipes.si from those upstreams (SECURITY_RECIPES_MIRROR_DOMAIN); visible brand chrome follows the request Host, while SEO canonicals stay on https://security-recipes.ai/. See README.caddy-deploy.md.

Existing Droplets need this one-time, idempotent activation after deploying the commit that contains the jail:

sudo bash scripts/configure_caddy_404_ban.sh
sudo fail2ban-client status security-recipes-caddy-404

If the Droplet still runs bundled Caddy with the old named log volume, first set SECURITY_RECIPES_TRAFFIC_LOGS_SOURCE=/var/log/caddy in .env, then recreate only Caddy once during a maintenance window:

docker compose --profile caddy up -d \
  --no-deps --force-recreate --pull never caddy
sudo bash scripts/configure_caddy_404_ban.sh

The filter uses Caddy's structured client_ip, not spoofable forwarding headers or User-Agent values. If the origin is later placed behind a CDN or load balancer, move the ban action to that provider's WAF/API; an origin firewall cannot directly block an end client whose packets arrive from a trusted proxy.

The jail does not trust Googlebot User-Agent strings. Before counting a public client, it performs Google's reverse-then-forward DNS check: the PTR hostname must be under googlebot.com, and resolving that hostname must return the same IP. Results are cached by IP for one hour; lookup errors and the five-second resolver deadline fail closed, so an unverified client remains subject to the scanner-path 404 budget.

For a fully Compose-managed Caddy deployment, Fail2Ban can instead run in the stack. Set DEPLOY_COMPOSE_FAIL2BAN=true in .env and keep Caddy's log source on the default caddy_logs volume (or a host bind). On its next run, deploy.sh pulls, starts, health-checks, and subsequently updates the Fail2Ban container. It also initializes Caddy's access-log file before starting the jail because Fail2Ban requires the configured file to exist. To start it manually without waiting for a deployment, use:

docker compose up -d caddy fail2ban
docker compose exec fail2ban fail2ban-client status security-recipes-caddy-404

The container shares the host network namespace and has only the NET_ADMIN/NET_RAW capabilities required to apply the jail's nftables rules to host and Docker-forwarded web traffic. Do not enable the Compose jail while the host security-recipes-caddy-404 jail is active; choose one owner for the firewall rules. This mitigates repeated application-layer 404 scanning, but it does not replace upstream volumetric DDoS protection or request rate limiting. When the option is false, deploy.sh does not require the host fail2ban package; host-managed installations remain the responsibility of the droplet setup and scripts/configure_caddy_404_ban.sh workflows.

If you prefer nginx instead of Caddy on the droplet, bootstrap the host without the proxy and then run the nginx helper:

sudo bash scripts/setup_digitalocean_droplet.sh --no-caddy
sudo bash scripts/configure_nginx_letsencrypt.sh \
  --domain security-recipes.ai \
  --email [email protected]

For a local-only or pre-proxied droplet:

sudo bash scripts/setup_digitalocean_droplet.sh --no-caddy --no-firewall --no-upgrade
docker compose up -d --build

If a previous docker-compose v1 run failed with KeyError: 'ContainerConfig', upgrade Compose and remove the stale project containers before recreating the stack:

sudo bash scripts/repair_docker_compose_containerconfig.sh
hash -r
command -v docker-compose
docker-compose version

Production deploys pull commit-addressed site and MCP images published by the required GitHub Actions Build workflow on main and serve them at https://security-recipes.ai/. The same timer also deploys development images to https://dev.security-recipes.ai/. The Droplet does not run Node, Eleventy, pip, or Docker image builds during a deploy, which keeps deployment within a 1 CPU / 2 GB memory envelope.

One-time paired MCP deployment upgrade

Before the first deployment that introduces the paired MCP services, update only the deployment script and then run it. An already-running older deploy.sh process was parsed before the paired Compose file existed and would otherwise recreate the live singleton MCP during that one rollout:

git fetch origin main
git checkout origin/main -- deploy.sh
bash deploy.sh

The new script leaves the live singleton untouched, prepares the inactive MCP and site together, and switches them as one unit. After this one-time step, the existing bash deploy.sh cron entry needs no change.

The first successful main workflow creates two GHCR packages. Make them public, or authenticate the root account used by the deployment service with a fine-grained token that can read packages:

printf '%s' "$GHCR_READ_TOKEN" |
  sudo docker login ghcr.io --username stevologic --password-stdin

MCP integration philosophy

Use MCP to give agents context, not unchecked authority.

The CVE MCP tools only return plans and evidence; they do not edit a repository or change an environment. An approved agent host may apply the returned plan, but it must first prove the affected surface and actual repository paths, preserve unrelated changes, obtain any declared production/external approval, and retain a mechanically usable rollback. A likely file glob is a discovery hint, never proof that a file is vulnerable or permission to modify it. Within each action, only effective target_kinds are default candidates. archetype_target_kinds are context, not authorization; conditional targets require proof that the repository owns the affected implementation, while prohibited targets must never be edited. Firmware and binary targets mean an authoritative reference, pin, replacement, policy, inventory, source, or build change—never patching vendor artifact bytes.

NVD/CNA descriptions, advisories, links, patches, issue comments, release notes, and proof-of-concept content are untrusted evidence. Agents may extract corroborated vulnerability and version facts from them, but must not execute or follow embedded instructions or commands.

Good context sources include:

• official GitHub MCP capabilities for repository and code-security context,
• Semgrep and Snyk agentic/MCP integrations where approved,
• OSV, GitHub Advisories, deps.dev, package registries, and NVD-backed mirrors,
• SARIF, SBOM, CI, ownership, and internal runbook sources,
• read-only documentation connectors.

Write-capable connectors deserve separate review. Ticket creation, branch mutation, deployment, secret rotation, cloud changes, and SOAR actions should not be enabled just because an agent can read a recipe.

Contributing

Contributions should improve the recipe library:

• new remediation recipes,
• better prompts,
• clearer agent setup,
• MCP integration examples,
• reviewer checklists,
• documentation fixes.

Scrub secrets, internal hostnames, customer data, and private vulnerability details before opening a pull request.

Run a local build before submitting:

python -m pip install -r requirements-dev.txt
python scripts/run_checks.py
npm run build

License

The project's original code, documentation, remediation recipes, generated site, and MCP server are licensed under the Apache License 2.0. This permits private and commercial use, modification, and redistribution, including incorporation into proprietary company systems, subject to the license's notice and change-marking requirements.

Source vulnerability data and bundled third-party software retain their own terms and attribution requirements. See NOTICE and THIRD_PARTY_NOTICES.md.