Stack And Dns Mcp

stack-and-dns-mcp

A monetized Model Context Protocol server that inspects any public domain for AI agents. Give it a domain and it returns the company's web technology stack, its email deliverability and DNS health (MX / SPF / DKIM / DMARC / MTA-STS / TLS-RPT / BIMI), and its HTTPS security posture, all as strictly-typed JSON with 0-100 scores and remediation-ready findings.

Built for Cursor, Claude Desktop, Claude Code, autonomous SDR bots and research agents. Runs locally over stdio with zero configuration, or as a hosted Streamable HTTP / SSE service that meters usage per API key and can sell prepaid credits through HTTP 402 payment challenges (Stripe hosted checkout or x402 / Machine Payments Protocol stablecoin payments) or API key subscriptions through a self-serve signup page.

Every scan is remembered, so it can also tell you what changed: domain_history shows when a company adopted or dropped a tool or weakened its DMARC policy, bulk checks work through whole lead lists (as MCP tools or a CSV-friendly REST API), watched domains raise alerts, by poll or signed webhook, when their email posture or stack changes, and find_domains_using answers "who runs Klaviyo?" from a robots.txt-respecting crawl of a domain list you supply.

Contents

• What the tools return
• Quick start (local, stdio)
• Client configuration
• Tool reference
• Hosted mode (HTTP) and monetization
• Configuration reference
• Defensive design
• Scoring model
• Signature database
• Deployment
• Publishing to MCP marketplaces
• Development and testing
• Troubleshooting

What the tools return

Every successful result is delivered twice: as pretty-printed JSON in content[0].text (for clients that only read text) and as structuredContent validated against a Zod output schema. Failures never crash the server; they come back as isError results with no structuredContent, carrying { "error": { "code", "message", "retryable", "details" } } in content[0].text and in _meta.error. Arguments that do not match a tool's input schema (an empty domain, a wrong type, a missing field) and unknown tool names are rejected by the MCP SDK before the tool runs, with a plain-text isError result (MCP error -32602: ...) that has no error object and is not charged; fix the arguments rather than retrying unchanged.

Example (abbreviated) comprehensive_audit highlights for a real store:

Platform: Shopify on Cloudflare (10 technologies detected).Analytics & tags: Elevar, Google Tag Manager, RudderStack.Go-to-market stack: Attentive.Security: B (81/100); TLS valid, HSTS on, CSP off.Email: Google Workspace; SPF fail; DKIM found; DMARC p=reject (score 92/100, A).

Quick start (local, stdio)

Requirements: Node.js 22.13+. Stored scans, the shared credit ledger and every feature built on them (domain_history, bulk checks, monitoring, reverse lookup) use the built-in node:sqlite module. When it cannot be loaded, or the database cannot be opened, the server still starts with the live inspection tools only and logs why.

git clone https://github.com/Mewnah/stack-and-dns-mcp.gitcd stack-and-dns-mcpnpm installnpm run build

Try it without any MCP client:

npm run smoke -- shopify.com            # comprehensive auditnpm run smoke -- shopify.com --tech     # tech stack onlynpm run smoke -- shopify.com --email    # email infrastructure only

Open it in the MCP Inspector:

npm run inspector

Or run the server directly (stdio is the default transport):

node dist/index.js

Logs always go to stderr as JSON lines; stdout is reserved for the MCP protocol.

Client configuration

The usage file and the database default to ./data/usage.json and ./data/stackdns.db, relative to the directory the server is started in. MCP clients usually start servers somewhere else (an application folder, /, whichever project is open), so the examples below set both to absolute paths. Otherwise the data lands in a different place per client, or storage is switched off with only a log warning when that directory is not writable.

Cursor

A project-level file is included at .cursor/mcp.json. Cursor runs project servers from the workspace root, so the relative paths work after npm run build. For a global install use absolute paths, with a checkout or with the published package:

{  "mcpServers": {    "stack-and-dns": {      "command": "node",      "args": ["/absolute/path/to/stack-and-dns-mcp/dist/index.js"],      "env": {        "STACKDNS_LOG_LEVEL": "warn",        "STACKDNS_USAGE_FILE": "/absolute/path/to/stack-and-dns-mcp/data/usage.json",        "STACKDNS_DB_FILE": "/absolute/path/to/stack-and-dns-mcp/data/stackdns.db"      }    }  }}

{  "mcpServers": {    "stack-and-dns": {      "command": "npx",      "args": ["-y", "stack-and-dns-mcp"],      "env": {        "STACKDNS_LOG_LEVEL": "warn",        "STACKDNS_USAGE_FILE": "/absolute/path/to/stack-and-dns-data/usage.json",        "STACKDNS_DB_FILE": "/absolute/path/to/stack-and-dns-data/stackdns.db"      }    }  }}

Claude Desktop

Edit claude_desktop_config.json (%APPDATA%\Claude\claude_desktop_config.json on Windows, ~/Library/Application Support/Claude/claude_desktop_config.json on macOS). A Windows example is in examples/claude_desktop_config.json.

{  "mcpServers": {    "stack-and-dns": {      "command": "node",      "args": ["D:\\Projects\\MCP\\stack-and-dns-mcp\\dist\\index.js"],      "env": {        "STACKDNS_LOG_LEVEL": "warn",        "STACKDNS_USAGE_FILE": "D:\\Projects\\MCP\\stack-and-dns-mcp\\data\\usage.json",        "STACKDNS_DB_FILE": "D:\\Projects\\MCP\\stack-and-dns-mcp\\data\\stackdns.db"      }    }  }}

macOS / Linux:

{  "mcpServers": {    "stack-and-dns": {      "command": "node",      "args": ["/Users/you/stack-and-dns-mcp/dist/index.js"],      "env": {        "STACKDNS_LOG_LEVEL": "warn",        "STACKDNS_USAGE_FILE": "/Users/you/stack-and-dns-mcp/data/usage.json",        "STACKDNS_DB_FILE": "/Users/you/stack-and-dns-mcp/data/stackdns.db"      }    }  }}

Restart Claude Desktop; the tools appear under the tools icon: the three inspection tools always, plus domain_history, the four bulk-check tools and the four monitoring tools when storage is available (12 in all), and find_domains_using / list_technologies when a crawl seed file is set.

Claude Code

claude mcp add stack-and-dns \  -e STACKDNS_USAGE_FILE=/absolute/path/to/stack-and-dns-mcp/data/usage.json \  -e STACKDNS_DB_FILE=/absolute/path/to/stack-and-dns-mcp/data/stackdns.db \  -- node /absolute/path/to/stack-and-dns-mcp/dist/index.js

Remote deployment:

claude mcp add --transport http stack-and-dns https://mcp.example.com/mcp --header "Authorization: Bearer sdm_live_..."

Metering a local stdio install with a customer key

A stdio server never contacts a hosted deployment; to use a hosted server's key, connect over HTTP as in the remote example above. STACKDNS_API_KEY only selects one of the keys configured in the same process (STACKDNS_API_KEYS or STACKDNS_API_KEYS_FILE in the same env block), so a local install can be metered and rate-limited under that key's plan. Without such keys the variable is ignored (the server logs a warning at startup) and the server runs in open local mode, unless STACKDNS_REQUIRE_AUTH=true, in which case every tool call answers UNAUTHORIZED. With keys configured but STACKDNS_API_KEY unset, or set to a key that matches none (that case logs a warning), the process runs as an anonymous free caller: the other tools work within the free plan's quota, and the bulk-check and monitoring tools answer UNAUTHORIZED.

Tool reference

The tools that take a domain (the three inspection tools, domain_history and watch_domain) accept it as a bare hostname (example.com), a URL (https://www.example.com/pricing) or an email address ([email protected]); start_bulk_check takes a list of the same forms. Scheme, path, port, query and a leading www. (for email checks) are stripped; IDN names are punycoded. An empty or whitespace-only domain fails the input schema (see above), not INVALID_DOMAIN.

inspect_tech_stack

{ "domain": "allbirds.com", "includeRawHeaders": false, "includeEvidence": true }

Key output fields:

verify_email_infrastructure

{ "domain": "example.com", "dkimSelectors": ["mycustomselector"], "checkMtaSts": true }

comprehensive_audit

{ "domain": "hubspot.com", "dkimSelectors": [], "includeRawHeaders": false, "includeEvidence": false }

Returns scores (overall, grade, email, security, infrastructure, weights and the infrastructure breakdown), the full techStack and email results (or techStackError / emailError if one side failed), merged findings, findingCounts, agent-friendly highlights[] and a summary sentence. The summary sentence's counts leave out the organizational record's DMARC findings when a subdomain's own _dmarc lookup failed; findingCounts counts every finding.

domain_history

{ "domain": "example.com", "changesLimit": 50 }

Reads stored scans only; it never contacts the domain. Every completed scan (a call to one of the three inspection tools, a bulk-check item, a scheduled monitor check or a reverse-lookup crawl) is recorded in the database named by STACKDNS_DB_FILE, keyed by the site with any leading www. removed. A call that was cancelled or hit its deadline is not recorded, because its lookups may have been cut short. Stored scans never record which caller requested them.

The first usable scan of a domain is a baseline and records no changes. A technology or DKIM selector must be missing from two consecutive clean scans before it is reported as removed, so a lazily loaded tag does not flap. Website scans that were blocked or failed are kept for the record but never change the stored state.

Email components are tracked separately: the MX records, whether the exchanges resolve, the SPF record, the SPF include tree, DMARC, MTA-STS, TLS-RPT/BIMI, and each DKIM selector. Each one updates, and is compared, only when its own lookups answered:

• A component whose lookup failed (SERVFAIL, a timeout, a policy host that answered 5xx or 408, a bot challenge) keeps its last known value and reports no change. The components that did answer in the same scan are still compared, so a failed include never hides a +all in the domain's own record.
• A definitive answer is a real change: an MTA-STS policy that returns 404, a policy host that does not exist (confirmed with our own resolver), or an invalid certificate.
• A failure that persists for three scans spanning at least a day is accepted as the answer, because receivers see the same failure: exchanges that never resolve become MX_ALL_UNRESOLVABLE, a policy host that never answers removes the MTA-STS mode. Nothing stays frozen forever. The streak only grows while the same lookup keeps failing, so a single timeout of something else is still held.
• The score only moves when every lookup answered or its failure was accepted as persistent; only a scan whose lookups all answered is stored with ok: true. While a subdomain's own _dmarc lookup fails, the stored score keeps the subdomain's last known DMARC points (the live score scores the unknown policy like p=none), so a timeout neither raises nor lowers it.
• DKIM risks and the score are compared only between scans that probed the same selectors. A key under a custom selector that one call passed and a bulk scan did not probe never reads as added or lost. Every email scan (calls, bulk items, scheduled checks) also probes every selector already known for the domain.
• Records are evaluated in a fixed order, so a domain that publishes two SPF or DMARC records does not flip policy between scans because resolvers return them in a different order.
• A subdomain whose own _dmarc lookup fails is never credited with the organizational domain's policy: the policy reads unknown (the live score scores it like p=none), and the organizational record's findings are not taken as the subdomain's (the one-line summary does not count them either; riskCounts still does). The subdomain's own DMARC stays in unverified until its own lookup has answered once, and that first answer is a baseline, not a change. A subdomain that simply has no record of its own still inherits the organizational policy. When the lookup of an SPF redirect= target fails, the stored and bulk policy reads unknown too, not none, and the findings that describe the terminal all (SPF_PASS_ALL, SPF_NEUTRAL_ALL, SPF_SOFTFAIL_ALL) keep their last known state. The live verify_email_infrastructure field spf.allPolicy still reads none in that case (its enum has no unknown); SPF_DNS_ERROR, the score note and the summary (SPF policy unknown (redirect=<target> lookup failed)) flag it.
• SPF whose redirect= chain has never answered whole (the target or a later redirect= hop failed) stays in unverified (with spfLookups) only until that failure is accepted as persistent. The domain's own record did answer, so from then on it is compared as usual while its policy reads unknown: a second record raises SPF_MULTIPLE_RECORDS; deleting the record, or an accepted failure of its own TXT lookup, is spf_removed; and a record that stops handing its policy to that target (another redirect= target, or an all of its own, which overrides redirect=) is compared like any other change. A deletion, an accepted lookup failure and a record that stops delegating each end the wait for the target, so whatever follows is compared as usual too. A second record ends it only when the record evaluated (the first in sorted order) no longer delegates to the target. Only what the redirect= chain holds was never read whole. While the record still delegates to the target, the first scan in which the whole chain answers is the baseline of the terminal policy and the findings that describe it (no spf_policy, risk_new or risk_resolved for them). The first such scan in which spfLookups is established too (the include tree answers, or an include's failure has been accepted) is the include tree's baseline: no spf_broken, spf_fixed, risk_new or risk_resolved for validity, lookup counts or void lookups (SPF_VOID_LOOKUPS, SPF_TOO_MANY_LOOKUPS, SPF_CIRCULAR_INCLUDE). That is a later scan than the policy's when an include lookup failed in the policy's scan, and it still reports a policy change made in between. Until the whole chain answers, the first scan in which the target answered but a later hop's failure was accepted is a baseline for the tree read so far (if the target already answered in the scan that took SPF out of unverified, that scan was this baseline); later such scans are compared with it. Each of these baselines covers the score too: no score_drop or score_rise in its scan, even for a score change another component caused. Findings of the domain's own record (such as SPF_MULTIPLE_RECORDS or SPF_SYNTAX_ERROR) and changes to other components are still reported in that scan.

Weakened policies (DMARC reject to none, SPF -all to ~all) are high, and an MTA-STS downgrade is medium; strengthened ones are info. An unknown policy ranks with none: reject or -all becoming unknown is high, and none becoming unknown is info. A policy read again after an unknown stretch is ranked against the policy last actually read before it, not against unknown: the same policy coming back (even +all or a malformed DMARC record) is info, while quarantine coming back as none, or -all as +all, is high. A move away from unknown with no policy read before it is info. In email, score is null until a scan in which every lookup answered or its failure was accepted as persistent. unverified lists components not yet established: no lookup has answered them and their failure has not been accepted as persistent (DMARC and SPF behind redirect= follow the rules above). lookupsFailing lists those whose latest lookups failed (their fields are the last known values, or the current reading once the failure was accepted as persistent), so an empty unverified does not mean every lookup answered: an accepted failure is listed in lookupsFailing. States stored by earlier versions are re-baselined silently on their next scan, with no change records: an organizational DMARC policy or its findings stored for a subdomain whose own lookup failure had been accepted, an SPF policy stored as none behind an accepted failed redirect= target, and a stored score from before SPF void-lookup and circular-include permerrors cost points, when the findings stored with that score and the scan that next rewrites it both carry such a permerror (a permerror that first appears after the upgrade is reported as a score_drop as usual). When a scan sets a DMARC or SPF baseline but cannot rewrite the score because another lookup failed, the next scan that rewrites the score re-baselines it silently.

Bulk checks: start_bulk_check, get_bulk_check, list_bulk_checks, cancel_bulk_check

{ "check": "email", "domains": ["acme.com", "https://www.globex.com/about", "[email protected]"] }

check is email (verify_email_infrastructure), tech (inspect_tech_stack) or audit (comprehensive_audit). The job is queued at once and worked through in the background by a small pool shared fairly by every job, so one large list cannot starve a small one; jobs survive restarts. start_bulk_check returns the job id, the inputs that were skipped (invalid, private, or a duplicate of an earlier site) and what was charged. Poll get_bulk_check with the job id: it returns status, per-status counts and one page of results (offset, limit up to 500, optional status filter such as failed). Each result is a compact summary: score, grade, MX provider, SPF and DMARC policy, DKIM found and high risks for email; reachability, HTTP status, security grade, CMS, e-commerce, CDN and technology names for tech; the four scores and top highlights for audit. In email summaries the SPF and DMARC policy can also read missing (no record) or unknown (the lookup failed, a redirect= target's lookup failed, or a subdomain's own _dmarc lookup failed), the DMARC policy invalid (a record without a valid policy), and the MX provider unknown (the MX lookup failed) or null MX (no mail). Jobs are private to the caller that submitted them, so bulk checks (and monitoring) need an identified caller (an API key, a prepaid session, or a stdio server's own local identity): anonymous callers on a server with STACKDNS_REQUIRE_AUTH=false share an address-based identity and are refused.

Billing happens once, when the job is accepted: prepaid callers pay the single-domain price per accepted domain (email 1, tech 1, audit 2 credits), and plan keys spend one unit of their daily quota per domain, all or nothing. Items that fail for a retryable reason (a timeout, a resolver outage) and items cancelled before they ran are refunded when the job finishes. A definitive answer such as DOMAIN_NOT_FOUND is a completed check and is charged. The polling, listing and cancelling tools are free and never touch the daily quota. Limits: STACKDNS_BULK_MAX_DOMAINS per job (default 1000) and STACKDNS_BULK_MAX_ACTIVE_JOBS unfinished jobs per caller (default 3). A list may carry up to twice STACKDNS_BULK_MAX_DOMAINS raw entries; a longer list, or one with more distinct domains than the maximum, is refused with LIMIT_EXCEEDED (details.maxDomains), over MCP as over REST. Only a list over max(50000, 2 x STACKDNS_BULK_MAX_DOMAINS) entries fails start_bulk_check's input schema first, which gives the plain-text -32602 result. A list with more invalid or duplicate entries than max(100, accepted domains) is refused with INVALID_ARGUMENT, since skipped entries are stored too. A prepaid session whose balance a job spent to zero can still read and cancel its jobs, even when its client keeps re-sending the payment header it paid with; only paid calls and new job submissions (POST /v1/jobs or start_bulk_check) answer with a new payment challenge.

Monitoring: watch_domain, get_alerts, list_watches, unwatch_domain

{ "domain": "acme.com", "checks": ["email", "tech"], "webhookUrl": "https://hooks.example.com/stackdns" }

A watched domain is re-checked every STACKDNS_MONITOR_INTERVAL_HOURS (default 24), once per domain however many callers watch it. A check that fails for a retryable reason (a timeout, a resolver outage) is retried after 30 minutes (or the interval, if shorter) and its prepaid charge is refunded; a definitive answer such as NXDOMAIN is charged, raises one domain_not_found alert and waits a full interval. Every change the scan store records (see domain_history) becomes an alert for each watcher: DMARC or SPF weakened, a DKIM key lost, the MX provider changed, a technology added or removed, a score drop, or the domain disappearing from DNS. Changes found by any scan reach watchers, including a direct call or a bulk job. A new watch only hears about changes from then on. Unwatching and watching again keeps the existing schedule, so it never forces an early check.

get_alerts returns unread alerts oldest first and marks them read (markRead: false to peek, sinceId and nextSinceId to page). With a webhookUrl (https only, a fully qualified public host name or public address, no redirects followed) each batch is also POSTed as JSON:

{ "type": "domain.changes", "watchId": "watch_...", "domain": "acme.com", "sentAt": "...",  "alerts": [{ "id": 42, "kind": "email", "change": "dmarc_policy", "subject": "dmarc", "oldValue": "reject", "newValue": "none", "severity": "high", "at": "..." }] }

The request carries x-stackdns-signature: t=<unix seconds>,v1=<hex>, where v1 is the HMAC-SHA256 of <t>.<raw body> keyed with the webhookSecret returned once when the watch is created. The timestamp is taken when the request is sent, so verify it and reject timestamps older than a few minutes. A delivery that fails is retried after 1 minute, 5 minutes, 30 minutes, 2 hours and 12 hours, then marked failed; the alert stays readable through get_alerts either way, with a short reason ("timed out", "DNS lookup failed", "HTTP 503"). Deliveries run beside the scheduled checks, at most one request at a time per account, so a slow endpoint delays only its own owner's alerts, and accounts take turns watch by watch within each delivery round, so neither one account's backlog nor a few slow receivers hold up everyone else's webhooks.

Plans include a number of watched domains (free 5, pro 100, enterprise 2000, payg 500, override with STACKDNS_MONITOR_WATCH_LIMITS_JSON). Prepaid callers pay STACKDNS_MONITOR_CREDITS_PER_CHECK (default 1) per scheduled check. When the balance runs out the watch pauses with a watch_paused alert, and it resumes by itself once the balance covers a check again. Self-serve accounts' watches follow the subscription: they pause (plan_inactive) while it is cancelled or unpaid, the newest beyond a downgraded plan's limit pause (over_plan_limit), and they resume when the plan allows them again. When the subscription is deleted for good (including after its key was already suspended for failed payments), its watches are removed, up to 200 per round across all such accounts, since no key can manage them any more; until then they stay paused and their pending webhook deliveries are cancelled, even a delivery that is in flight at that moment. Watching, listing and reading alerts are free. Monitoring needs an identified caller (an API key, a prepaid session, or a stdio server's own local identity); watches that anonymous callers created on earlier versions are removed by the schema upgrade.

Reverse lookup: find_domains_using, list_technologies

{ "technology": "Klaviyo", "activeOnly": true, "seenWithinDays": 90, "limit": 50 }

Returns domains that run a technology, most recently seen first, with the detected version and first-seen and last-seen dates, plus nextCursor for the next page. technology is a signature id or name; an ambiguous name returns INVALID_ARGUMENT with the matching candidates in details.candidates. list_technologies browses the catalogue (filter by query or category) with how many domains currently use each technology.

The searchable universe is the crawl list, not the internet. An operator points STACKDNS_CRAWL_SEED_FILE at a file of domains (one per line, or the first column of a CSV, re-read when it changes), and a background crawler works through it at STACKDNS_CRAWL_RATE_PER_MINUTE (default 20), re-visiting each domain every STACKDNS_CRAWL_REFRESH_DAYS (default 30). For each domain it:

• reads https://<domain>/robots.txt first and skips the site when the group for our user agent (StackAndDnsBot, also matched when a group names it with a version such as StackAndDnsBot/1.0), or the * group when no group names it, does not allow / (RFC 9309: a group naming us overrides *, the rule with the most octets wins, rules and paths are compared in one percent-encoded form, so Disallow: /über/ also covers /%C3%BCber/, and a literal * or $ is named by its escape, %2A or %24);
• fetches only the homepage, with the same transparent User-Agent, 3-second per-attempt deadline (6 s at most) and SSRF guard as inspect_tech_stack, and checks every URL the fetch requests (a redirect, the http:// fallback) against the robots.txt of that URL's origin first; each origin's file is downloaded once per crawl, and the time spent reading it counts neither against the homepage's deadline nor against the time left for the insecure-TLS retry and the http:// fallback; when HTTPS fails and robots.txt refuses the http:// fallback, the domain is recorded as unreachable and its last error names both the HTTPS failure (its first 300 characters) and the refusal;
• records the result like any other scan, so crawled domains also build up domain_history and can raise monitoring alerts;
• treats a robots.txt that answers 5xx or 429, or cannot be fetched, as "not now": the site is skipped, keeps its data and is retried on the failure schedule (after 1 hour, doubling);
• waits a full refresh before retrying a domain that does not exist, and removes a site's data from the reverse lookup (and from universe.scanned) when its robots.txt opts out or the domain stops existing.

The reverse lookup answers only from what the crawler itself observed, for domains on the crawl list. Customers' calls, bulk jobs and monitor scans never appear there (not even their timing for a crawl-list domain), so the tool cannot reveal anyone's prospect list. A changed seed file is applied once it has stopped changing for one round, so a file being rewritten is never read half written, and in small batches with pauses between them, so even a list of millions of domains neither stalls the server nor holds up other processes' writes to the database. The list is parsed into a private in-memory SQLite database and compared with the crawl targets there. That database is attached only while the sync runs (about 200 MB for 5 million domains) and freed when it ends; the list is never held in the JavaScript heap. At most STACKDNS_CRAWL_MAX_TARGETS distinct valid domains are taken from the file, in file order. universe.domains and universe.scanned in each answer say how large the searchable set is; GET /healthz reports crawl progress. Without a seed file the two tools are not offered.

Error codes

These codes come from the tools themselves. Arguments that fail a tool's input schema (an empty or whitespace-only domain included) never reach it: the SDK answers with a plain-text MCP error -32602: ... result that carries no code, so such input never yields INVALID_DOMAIN.

Payment redemption over HTTP answers 503 payment_recording_failed when a payment verified (or settled) but could not be recorded yet; retrying the same credential is safe and never pays twice. For x402 the 503 carries X-Session-Token (and sessionToken in the body) for the principal the credits will land on: keep it. Any process sharing the database finishes the recording in the background, and until then that token is answered 503, never a new payment demand. A retry of the same X-PAYMENT header completes the recording too, but it never hands out a token again: a settled authorization is public on-chain, so presenting it proves nothing about who paid. The 503's principal gets nothing when the credits go to another presentation of the same authorization (with or without from) instead: one that settled earlier, even one first answered unconfirmed, or one a sibling process recorded first because it could not see this settlement (see Any Node host). A retry of that 503's header then answers 409 payment_already_redeemed.

When the facilitator does not confirm a settlement (a timeout or a 5xx on /settle, or success without a transaction id), the answer is 502 x402_settlement_unconfirmed, again with the session token (left out only when the server could not store the unconfirmed state; the operator then issues it with npm run settlements -- session <principalId>, the principal being in the server's log line), never a new 402: the transfer may have gone through, so do not pay again. A prepaid session that can serve the request is served from its balance instead, with X-Payment-Status: x402_settlement_unconfirmed. That means a session with credits left, or any session on a request that spends nothing, such as reading or cancelling jobs. The 502 is only for requests nothing else can serve. The server keeps that state as long as it remembers redeemed payments (400 days), across restarts and sibling processes; with STACKDNS_DB_FILE=none it is kept in the usage file. A retry of the same header settles the payment if the facilitator shows the authorization unspent, and otherwise answers "unconfirmed", never x402_expired. The facilitator is asked again at most once a minute per payment, whichever process gets the retry; retries in between are answered from the stored state. For 24 hours after the payment's last update (a re-check counts), the session it would credit answers a spending request its balance cannot cover with that 502 rather than a payment demand. That holds whether or not its client re-attaches the payment header. After that the session may buy again: a spending request it cannot cover gets the 402 credits_exhausted payment demand, and when the unconfirmed header is attached the 402 also carries X-Payment-Status: x402_settlement_unconfirmed. The header sent without a session keeps answering 502. Contact the operator with the header to reconcile. The server logs the payer, amount and target principal, and the operator resolves the payment with npm run settlements (see Reconciling x402 settlements).

A retry of a payment that is already recorded answers 409 payment_already_redeemed, even when its stored settlement state could not be removed at the time.

If the answer to a 502 or 503 is lost (for example a client timeout), retrying the payment without a session does not return the token: a settled authorization is public on-chain, so presenting it proves nothing. The operator can issue the session with npm run settlements -- session <principalId>.

One narrow case remains. Suppose the database fails both the 'settled' marker write and the recording. The 503 still goes out with the session, and the process then stops before its background retry publishes the marker. A later presenter of the same (public) payment header could then be handed that session, but only if the facilitator also settles an already-spent authorization again, or the transfer is still pending. All three must coincide. The error log line settled x402 payment answered with its session, but its 'settled' marker could not be written flags such a payment for manual reconciliation.

An unreachable or blocked homepage is not an error: the tool returns fetch.ok=false / blocked.blocked=true with the TLS and DNS data it could still gather.

Hosted mode (HTTP) and monetization

node dist/index.js --http --port 3333

Endpoints:

Rate-limit headers X-RateLimit-Limit, X-RateLimit-Http-Limit and X-RateLimit-Plan are set on every authenticated response, X-RateLimit-Remaining and X-RateLimit-Reset on each one within the HTTP request ceiling, and a 429 from that ceiling carries Retry-After instead; X-RateLimit-Limit is the plan's tool calls per minute and X-RateLimit-Http-Limit the ceiling on HTTP requests of any kind (initialize, tools/list, ping and tool calls). Requests that end before a principal is resolved (invalid keys or tokens, payment redemption, 402 challenge minting) are throttled per client address (STACKDNS_UNAUTHENTICATED_PER_MINUTE, default 60; set STACKDNS_TRUST_PROXY correctly behind a load balancer or every client shares one address). A prepaid session that can serve the request and re-attaches a payment header the server answers without asking Stripe or the facilitator (already redeemed, unconfirmed, expired, or still being settled or recorded) counts against its own ceiling instead. Sessions can only be driven or ended by the principal that initialized them. A request for another principal's session answers 404 so a client whose identity changed simply re-initializes. Per-address limits treat an IPv6 client's whole /64 as one address.

With STACKDNS_REQUIRE_AUTH=false, requests without credentials run as an anonymous principal per client address, on the free plan whenever the server sells access (keys, payments or signup) and on local only on a server that sells nothing. Such requests from a browser page on an origin not listed in STACKDNS_ALLOWED_ORIGINS are refused with 403, so a web page cannot act as its visitor.

API keys

Generate keys (the raw key is shown once; the server stores only the SHA-256 hash):

npm run keys:generate -- --plan pro --label "Acme Corp"npm run keys:generate -- --plan enterprise --count 5 --json > new-keys.json

The --json output contains the raw keys. Hand them out, then keep only id, keyHash, prefix, plan, label and active in the keys file and delete the raw output; the plain output prints that hashes-only block for you. A key file that still has key fields works, but leaves the raw keys on the server. .gitignore covers keys.json and new-keys.json at the project root so neither is committed by accident.

Load them either inline or from a file:

# small deploymentsexport STACKDNS_API_KEYS="sdm_live_abc...:pro:Acme Corp,sdm_live_def...:free:Trial"
# recommended: hashes only, no raw keys on the serverexport STACKDNS_API_KEYS_FILE=/etc/stackdns/keys.json

keys.json:

[  { "id": "key_1f2e3d4c5b6a", "keyHash": "<sha256 hex>", "prefix": "sdm_live_abc12", "plan": "pro", "label": "Acme Corp", "active": true }]

Clients send Authorization: Bearer sdm_live_... or X-API-Key: sdm_live_.... A key file path that does not exist yet (a fresh volume, say) is logged and treated as zero file-backed keys; a file that exists but cannot be parsed is a startup error.

Self-serve API keys (Stripe subscriptions)

Set STRIPE_SECRET_KEY and a recurring Stripe Price for one or both plans (STRIPE_PRICE_PRO=price_..., STRIPE_PRICE_ENTERPRISE=price_...) and GET /signup becomes a plan page. The buyer subscribes through Stripe Checkout and lands on /keys/success, which issues the key and shows it exactly once; a second visit (the Checkout Session id in the URL is a payment reference that leaks through logs and dashboards) never shows or re-issues it. A payment that is still clearing (bank debits) gets its key on that first visit too, inactive until Stripe confirms the payment, so the id cannot be used by someone else to claim the key first. A subscription that has already ended when the buyer opens the page (Stripe reports it canceled or expired, or a subscription webhook has recorded its cancellation, expiry or deletion) answers 410 and shows no key; so does a reload of that page. The key's plan follows the subscription's current price. Keys are stored as SHA-256 hashes in the database next to the configured ones and work on every route. A buyer replaces a key with POST /v1/keys/rotate (authenticated with the current key; available while signup is configured, and after it is switched off the operator rotates keys), and the account keeps its bulk checks, watches, alerts and usage across rotations.

Point a Stripe webhook at POST /v1/stripe/webhook with the events customer.subscription.created, customer.subscription.updated and customer.subscription.deleted, and set STRIPE_WEBHOOK_SECRET=whsec_.... Keys are then suspended while a subscription is past due, unpaid or paused, restored when it recovers, moved to the new plan when the price changes, and revoked for good when the subscription is deleted. Events are applied once and in order, and signatures older than five minutes are refused. The webhook stays mounted whenever STRIPE_WEBHOOK_SECRET is set, even after signup is switched off, so keys already sold keep following their subscriptions. Without the webhook secret, keys keep working after a cancellation, and the server warns about it at startup. Signup and checkout requests share the per-address ceiling for unauthenticated traffic.

Bulk checks over REST

The bulk-check tools are also a plain REST API for spreadsheets and scripts, with the same billing, limits and ownership:

# CSV or plain text: the first column of each line (a header row such as "domain" or "website" is skipped)curl -X POST "https://mcp.example.com/v1/jobs?check=email" \  -H "Authorization: Bearer sdm_live_..." -H "Content-Type: text/csv" --data-binary @leads.csv
# or JSONcurl -X POST https://mcp.example.com/v1/jobs -H "Authorization: Bearer sdm_live_..." \  -H "Content-Type: application/json" -d '{"check":"tech","domains":["acme.com","globex.com"]}'
curl -H "Authorization: Bearer sdm_live_..." https://mcp.example.com/v1/jobs/job_...            # status and one page (?offset, ?limit, ?status)curl -H "Authorization: Bearer sdm_live_..." https://mcp.example.com/v1/jobs/job_.../results.csv -o results.csvcurl -X DELETE -H "Authorization: Bearer sdm_live_..." https://mcp.example.com/v1/jobs/job_...  # cancel pending items

POST answers 202 with the job and a Location header. The CSV export has one row per submitted line in the original order, with the check's summary columns; cells that a spreadsheet would run as a formula are prefixed with '. GET /v1/jobs/:id refuses an unknown ?status and a non-numeric ?offset or ?limit with 400 instead of ignoring them.

Errors from the job service, and request bodies that are not valid JSON or are too large, use the tool error shape { "error": { "code", "message", "retryable", "details" } } with a matching HTTP status (400 invalid input, 401 anonymous caller, 402 insufficient credits, 404 unknown job or route, 413 too many domains or a body over the limit, 429 quota or too many active jobs). A JSON body may be 1 MiB and a CSV body 2 MiB, or more when STACKDNS_BULK_MAX_DOMAINS is raised (64 bytes for each of the twice-the-maximum raw entries a job accepts). Requests turned away by authentication before the route runs (401 bad key or token, 403 browser origin, 409, 429 request ceiling, 502, 503) use { "error": "<code>", "message": ... } like every other HTTP route, and a 402 carries the payment challenge body.

Plans and rate limits

The daily cap of payg and local cannot be reached at their stock per-minute rates; it only matters when an override raises perMinute. Override any value with STACKDNS_RATE_LIMITS_JSON='{"free":{"perMinute":5},"pro":{"perDay":20000,"httpPerMinute":600}}' (httpPerMinute defaults to five times perMinute, never below 30). All four values are reported by GET /pricing. Usage is tracked per principal and per tool and persisted atomically to STACKDNS_USAGE_FILE; idle per-address and per-payment principals with no remaining credits are dropped from the ledger after 90 days, API-key counters are kept forever. With STACKDNS_DB_FILE=none, unfinished x402 settlements are kept in the usage file too (format version 3; older files still load), so a restart still knows a payment whose settle never answered. Pointing the server at a database later moves them into it.

HTTP 402 payment challenges

Enable with STACKDNS_PAYMENTS_ENABLED=true and one or both methods in STACKDNS_PAYMENT_METHODS. With STACKDNS_REQUIRE_AUTH=true (the HTTP default), a request that carries no credentials, a prepaid session that has run out of credits on a paid call or a job submission, and a payment credential that was refused (unpaid, expired, rejected by the facilitator) all receive the following (an invalid key or session token gets 401, and an unparseable payment credential gets 400):

HTTP/1.1 402 Payment RequiredWWW-Authenticate: Payment id="chl_...", realm="stack-and-dns-mcp", methods="stripe x402",                  intent="charge", expires="2026-09-16T22:00:00.000Z", request="<base64url JSON>"Content-Type: application/json
{  "x402Version": 1,  "error": "payment_required",  "accepts": [ { "scheme": "exact", "network": "base", "maxAmountRequired": "500000", "asset": "0x8335...", "payTo": "0x...", ... },               { "scheme": "stripe", "extra": { "checkoutUrl": "https://checkout.stripe.com/...", "paymentIntentId": "pi_...", "clientSecret": "..." } } ],  "challenge": { "id": "chl_...", "methods": ["stripe", "x402"], "instructions": { ... } }}

The header follows the Machine Payments Protocol WWW-Authenticate: Payment shape and the body is x402-compatible, so both agent ecosystems can parse it.

Stripe flow

1. Set STRIPE_SECRET_KEY, optionally STRIPE_PRICE_CENTS (default 500 = $5.00), STRIPE_CREDITS_PER_PURCHASE (default 100) and STRIPE_CURRENCY.
2. The challenge carries a PaymentIntent clientSecret (for programmatic confirmation) and a hosted Checkout checkoutUrl (for a human operator).
3. After paying, the agent retries with Authorization: Payment <base64url({"method":"stripe","challengeId":"chl_...","paymentIntentId":"pi_..."})> (or checkoutSessionId from the success page).
4. The server verifies the PaymentIntent through the Stripe API (status, amount, currency, service metadata, not already redeemed), grants credits and answers the original request. The response carries Payment-Receipt (base64url JSON) and X-Session-Token.
5. Subsequent calls use Authorization: Bearer sdm_sess.... until the credits run out, at which point a fresh 402 is issued. Session tokens expire 30 days after issue (an expired token gets 401 invalid_session_token). Every successful payment or top-up returns a fresh token in X-Session-Token and the Payment-Receipt; switch to it, because an expired token can no longer reach the credits left on it.
6. To top up without changing identity (and therefore without losing an open MCP session), keep Authorization: Bearer <existing sdm_sess token> and add X-PAYMENT (or send Authorization: Payment <credential> plus X-Session-Token: <existing token>); the credits are added to that principal, the same token stays valid until its own expiry, and the response carries a fresh 30-day token. An invalid or expired token sent with a payment is rejected with 401 before anything is charged.
7. Re-presenting an already-redeemed payment credential answers 409 (never 402) and carries the current payment challenge (WWW-Authenticate header and accepts body), so a client that keeps re-sending its last payment header learns how to buy more without auto-paying on a duplicate. A credential that is mid-verification in a concurrent request answers a plain 409 payment_redemption_in_progress with Retry-After: 2. Neither is a 402, so a duplicate retry can never charge twice. A payment reference alone is deliberately not a credential: it never re-issues a session token, so keep the token from the Payment-Receipt. A funded prepaid session that also carries a stale or unredeemable payment header is served from its own balance (with an X-Payment-Status header naming the reason). This includes a payment whose settlement is unconfirmed (X-Payment-Status: x402_settlement_unconfirmed); it also is not counted against the per-address limit while its stored answer is used. So is an exhausted session on requests that cannot spend credits (reading or cancelling jobs). A session re-sending a payment header that is already redeemed is limited by its own plan's rate limit, not by the per-address limit for unauthenticated requests.
8. Redeemed payment ids are remembered for 400 days (well past card dispute windows) and a PaymentIntent older than that is refused with payment_too_old, so an id that has aged out of the ledger can never be presented again for fresh credits. An x402 authorization whose validBefore has passed is refused as x402_expired before any facilitator call for the same reason.

x402 / MPP flow

1. Set X402_FACILITATOR_URL (a facilitator exposing /verify and /settle), X402_PAY_TO (your wallet), and optionally X402_NETWORK, X402_ASSET, X402_ASSET_NAME, X402_ASSET_VERSION, X402_PRICE_ATOMIC, X402_CREDITS_PER_PAYMENT. X402_ASSET_NAME and X402_ASSET_VERSION are the token's EIP-712 domain, which clients sign over (extra.name and extra.version in every requirement; GET /pricing shows them as payments.x402.assetName and payments.x402.assetVersion), so they must equal the contract's name() and version(). X402_ASSET, X402_ASSET_NAME and X402_ASSET_VERSION each default to USDC on X402_NETWORK: on base (or eip155:8453) 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913, USD Coin, 2; on base-sepolia (or eip155:84532) 0x036CbD53842c5426634e7929541eC2318f3dCF7e, USDC, 2. On any other network, startup with x402 enabled fails unless X402_ASSET and X402_ASSET_NAME are set (the version still defaults to 2). With x402 enabled, startup also refuses the built-in USDC contract of one network on the other, a name or version that differs from that contract's EIP-712 domain, and an X402_ASSET that is a token other than the network's USDC without X402_ASSET_NAME, because every payment would fail; set X402_ASSET_VERSION too if that token's version() is not 2. X402_NETWORK is matched against the built-in names case-insensitively and sent in their canonical spelling (Base-Sepolia is sent as base-sepolia, EIP155:8453 as eip155:8453), because clients and facilitators compare network ids exactly; any other network id is sent as written, only trimmed, since CAIP-2 references can be case-sensitive.
2. The agent signs an exact payment for the advertised requirements and retries with the X-PAYMENT: <base64 payload> header (or Authorization: Payment ...).
3. The server calls the facilitator to verify and settle, records the transaction, qualified by the authorization's payer and nonce, to prevent replay, grants credits and returns the receipt and session token as above.

Set STACKDNS_SESSION_SECRET so session tokens survive restarts, and give every process that shares the database the same value: a token signed by one process is otherwise rejected (401) by the others, including the token a 503 or 502 hands out, which no retry re-issues. Credits are refunded automatically when a metered call fails.

Reconciling x402 settlements

npm run settlements -- list                      # payments that need an operator (oldest first, at most 1000)npm run settlements -- list --all --jsonnpm run settlements -- resolve <paymentId> --tx <transaction hash>npm run settlements -- resolve <paymentId> --failedSTACKDNS_SESSION_SECRET=... npm run settlements -- session <principalId>

The tool works on the database in STACKDNS_DB_FILE (or --db <file>) and is safe to run next to live servers: each change is one transaction that first checks the marker is still as listed. It does not edit markers kept in the usage file (STACKDNS_DB_FILE=none). From a build, run node dist/cli/settlements.js <command> instead; in the Docker image, for example, docker exec <container> node dist/cli/settlements.js list. Exit codes: 0 done, 1 refused or failed, 2 bad usage. Bad usage is an unknown command or option, a flag without its value, a missing or extra argument, resolve without exactly one of --tx and --failed, or a --limit that is not a positive whole number, all checked before the database is opened; a marker that needs --nonce and is not given a usable one exits 2 as well.

• list shows unconfirmed payments (unknown), settles whose process died (settling (stale)) and settled payments whose credits are not recorded (settled), oldest update first. Each row has the payment id, principal, credits, age, time since its last update and the transaction. --all adds live settles and leftovers of payments that are already recorded; --json prints JSON. At most 1000 rows are shown (--limit <n> changes that). When more remain, a notice on stderr says so; stdout stays the table or the JSON.
• resolve <paymentId> --tx <hash>: you confirmed the transfer on-chain. The payment is recorded under the ids a server uses, and its credits go to the principal the marker names, once, however often you run the command, unless another presentation of the same authorization was credited first (it is then recorded without credits and reported as already recorded). Unlike a server, resolve does not yield to an earlier settlement of the same authorization that is not recorded yet: when list shows two settled markers with the same transaction, resolve the one with the greater age (claimed first) first, or let a running server record them. A settled marker accepts only the transaction the facilitator reported, and ignores --nonce. Any other marker written by an earlier release does not record its authorization, so pass --nonce <nonce> too (the EIP-3009 nonce from the payer's X-PAYMENT header, or from the AuthorizationUsed event); the payment is then recorded under that nonce alone. It is not credited if a server already credited the same authorization with its payer, and it blocks any later credit of that authorization. For a newer marker, --nonce is optional and must match the authorization's nonce. A settle that is still running is refused.
• resolve <paymentId> --failed: you confirmed nothing moved. The marker ends and the payment id is recorded as redeemed, so that authorization can never settle into credits on this server later. A payer whose authorization is still valid can then pay again with a new authorization. The old authorization stays valid on-chain until its validBefore; EIP-3009 tokens such as USDC let the payer cancel an unused one with cancelAuthorization. Settled markers are refused, because money moved: use --tx.
• session <principalId> prints a prepaid session token for the principal on stdout, signed with STACKDNS_SESSION_SECRET (required: it must be the HTTP servers' secret). Use it for a payer whose answer was lost. The principal is the one in the payment's log line or in list. The label and balance come from the database when one is available (--label <text> overrides the label); the command also works without one.

Configuration reference

Every variable is optional, and the server reads only its environment. See .env.example for the annotated template: copy it to .env and load it with npm run start:env (node --env-file=.env dist/index.js) or docker run --env-file .env, or set the variables in your process manager. With npm run start:env, variables already set in the environment take precedence over the file; with docker run --env-file, the file overrides the image's own settings (and -e flags override the file). With docker run --env-file, leave STACKDNS_TRANSPORT, STACKDNS_USAGE_FILE and STACKDNS_DB_FILE out of the file (they are commented out in the template). The image sets http and the /data paths, and a relative ./data is not writable by the image's non-root user. A value that fails validation stops startup with a message naming the variable.

Command-line flags override the matching variables: --http (aliases --sse, --remote), --stdio, --host <h>, --port <n> and --log-level <level>, each value also accepted as --flag=value. A missing or invalid flag value stops startup too; unrecognised arguments are ignored.

Defensive design

• Strict input validation. Zod schemas gate every argument; the domain parser accepts URLs, emails and IDN names but rejects IP literals, single labels, malformed labels and reserved TLDs.
• SSRF protection. Every outbound socket (HTTP fetch and TLS probe) resolves through a guarded lookup that refuses loopback, RFC 1918, CGNAT, link-local, cloud-metadata, multicast, reserved, ULA, IPv4-mapped and NAT64-embedded private addresses at connect time, which also closes the DNS-rebinding window.
• Hard deadlines everywhere. Per-attempt fetch deadline (default 3 s; the TLS and http:// fallbacks end by twice that), per-hop redirect budget (max 5 hops, http/https only), TLS handshake timeout, c-ares timeout plus a hard Promise race on every DNS query, and a whole-call ceiling in the meter.
• Body caps and charset-aware decoding. Streams are cut at STACKDNS_MAX_BODY_BYTES; the result is flagged truncated.
• Graceful degradation. TLS validation failures trigger an insecure retry (flagged), an unreachable HTTPS origin triggers an HTTP fallback (flagged), WAF challenge pages are detected and reported as blocked, and an unreachable homepage still yields TLS and DNS data.
• Resolver fallback. When the host resolver refuses raw UDP (common with VPN clients and container stubs), queries are retried on the public resolvers and preferred there for a bounded window; the first successful primary answer ends the window early.
• DNS answer cache. Successful answers (including NXDOMAIN/NODATA) are reused for 30 seconds and identical in-flight queries are coalesced, so the existence probe, the MX and SPF evaluations and both halves of comprehensive_audit ask each question of the resolver once. Errors are never cached.
• Cancellation. A tool call that times out or is cancelled by the client aborts its in-flight fetch, TLS probe and policy download and stops fanning out further DNS probes, so abandoned work does not keep consuming the concurrency budget.
• No crashes. Tool exceptions become structured isError results, credits are refunded on failure, and uncaughtException / unhandledRejection are logged rather than fatal.
• Secrets hygiene. API keys are stored hashed and compared in constant time; session tokens are HMAC-signed; logger redacts fields that look like secrets.

Scoring model

Email deliverability (0-100): MX 20, SPF 22, DKIM 20, DMARC 28, MTA-STS 5, TLS-RPT 3, BIMI 2. +all, permerrors (multiple records: 3 of 22 SPF points; more than 10 lookups, more than 2 void lookups or a circular include: at most 6 of 22), SPF syntax errors (at most 8 of 22), weak or revoked DKIM keys, p=none and partial pct all reduce points; disposable domains are capped at 10; NXDOMAIN is 0. An MTA-STS DNS record whose policy file cannot be fetched or is not a valid policy (no version: STSv1) earns 0 of 5; a DMARC lookup that fails scores as absent but is reported as unconfirmed (DMARC_DNS_ERROR plus a medium DMARC_MISSING). A subdomain whose own _dmarc lookup fails while the organizational record answers scores its unknown policy like p=none (10 of 28); the stored score keeps the subdomain's last known DMARC points instead. Once exchanges that never resolve are accepted as unresolvable, the stored score takes their MX points away as it would for exchanges that do not exist.

Security posture (0-100): security-header score x 0.65 (HTTPS 15, HSTS 15, CSP 15, clickjacking 10, X-Content-Type-Options 10, Referrer-Policy 8, Permissions-Policy 7, COOP 4, CORP 3, version disclosure 5, cookie flags 8) plus TLS 35 (valid chain, hostname and dates; deductions for imminent expiry; 8 points for a reachable but invalid certificate). HSTS earns nothing on a page served over plaintext HTTP, where browsers must ignore it, and a present but empty, zero or unparseable HSTS header scores 3 with an HSTS_INVALID finding. Referrer-Policy is judged by its last recognised token: the leaky values (no-referrer-when-downgrade, unsafe-url) and unrecognised values score 3 with a REFERRER_POLICY_WEAK finding.

Infrastructure maturity (0-100): presence of edge protection 15, analytics 15, tag management 10, go-to-market stack 15, modern framework 10, recognized platform 10, consent management 8, observability 7, customer engagement 5, experimentation 5. Blocked fetches are reported as a lower bound.

Overall: email 40 % + security 35 % + infrastructure 25 %. Grades: A >= 90, B >= 80, C >= 70, D >= 60, otherwise F.

Signature database

src/signatures/signatures.json holds 450+ entries validated at startup by src/signatures/schema.ts. Each entry:

{  "id": "klaviyo",  "name": "Klaviyo",  "category": "email-marketing",  "website": "https://www.klaviyo.com",  "confidence": 100,  "html": ["static\\.klaviyo\\.com", "_learnq"],  "scripts": ["klaviyo\\.js"],  "headers": { "x-powered-by": "^Klaviyo" },  "cookies": ["__kla_id"],  "meta": { "generator": "^Klaviyo ?([\\d.]+)?\\;version:\\1" },  "implies": ["shopify"],  "excludes": []}

• Patterns are case-insensitive regular expressions; a trailing \;version:\N captures a version from group N (Wappalyzer convention).
• headers values are regexes; an empty string means "header present".
• implies adds transitive detections (e.g. WooCommerce implies WordPress) at reduced confidence.
• Invalid patterns are logged and skipped so one bad entry never disables detection.
• npm test asserts that every pattern compiles, ids are unique and categories exist.

Deployment

Docker

docker build -t stack-and-dns-mcp .openssl rand -hex 32 > stackdns.secret   # once; keep it outside the repository (*.secret is git-ignored)docker run -p 3333:3333 -v stackdns-data:/data \  -e STACKDNS_PUBLIC_URL=https://mcp.example.com \  -e STACKDNS_API_KEYS_FILE=/data/keys.json \  -e STACKDNS_SESSION_SECRET="$(cat stackdns.secret)" \  stack-and-dns-mcp

The session secret must stay the same across container re-creation and upgrades, or outstanding session tokens, and the prepaid credits reachable only through them, stop working. Passing it through --env-file works as well; keep the template's transport and data-path lines commented out (see Configuration reference).

The image runs as a non-root user, exposes /healthz for orchestrators and keeps usage and the SQLite database (STACKDNS_DB_FILE=/data/stackdns.db) on the /data volume. Without a named volume or bind mount for /data, the data (stored scans, change history, prepaid balances) lives in an anonymous volume tied to that one container. It survives a restart, but a re-created container (an upgrade, a new docker run) starts empty, and --rm deletes it. The healthcheck probes /healthz on the port from PORT or STACKDNS_PORT (default 3333) and passes without probing when STACKDNS_TRANSPORT=stdio, which has no HTTP listener. A transport or port given only as a command-line flag (--stdio, --port) is invisible to it, so set the variables instead, or run with --no-healthcheck.

Any Node host (Fly.io, Render, Railway, ECS, a VM)

Set STACKDNS_TRANSPORT=http, PORT, STACKDNS_PUBLIC_URL, a fixed STACKDNS_SESSION_SECRET (the same for every process), STACKDNS_TRUST_PROXY=true when behind a load balancer, and mount persistent storage for STACKDNS_USAGE_FILE and STACKDNS_DB_FILE (an ephemeral disk silently resets stored scans, change history and balances on every deploy). Start with node dist/index.js --http. Streamable HTTP sessions and rate-limit windows are held per process, so run a single replica (or pin clients to one instance with sticky sessions and accept per-instance quotas). Overlapping processes that share the database file are safe:

• Prepaid balances and redeemed payment ids live in the database, so a charge in one process and a refund in another land on the same balance, and a payment is redeemed exactly once. A usage file from before this is imported once.
• A lease in the database elects a single worker, so during a rolling deploy, or with several stdio servers started by one client, bulk items, monitor checks and crawls run once.
• Every bulk item write is fenced to the worker that claimed it, so a worker that lost the lease can never overwrite its successor's results.
• Each scheduled monitor check is claimed in a transaction that re-checks it is still due, so overlapping rounds from two processes run and charge it once.
• A process that finds the database locked at startup (a sibling's migration, say) waits for up to two minutes, then exits rather than run on a private ledger beside the shared one. A stdio server, which sells nothing, waits up to 15 seconds to open the database and again to attach the shared credit ledger, and otherwise runs without storage, so MCP clients do not time out its initialize. An HTTP server waits up to 15 seconds for that first attach before it exits.
• Every process must share one STACKDNS_SESSION_SECRET; otherwise a prepaid session token issued by one process is rejected (401) by the others.
• x402 settlements in progress are recorded in the database before the facilitator is asked to settle, so a retry that reaches a sibling process, or arrives after a restart, never gets a new payment demand for money that may already have moved.
• Each settled x402 payment is recorded under its transaction qualified by its authorization (x402_<tx>:<payer>.<nonce>; the payer is the authorization's from, or the payer the facilitator's /verify reports when from is left out, given as a 20-byte hex address or as a CAIP-10 account id such as eip155:8453:0x...). A facilitator that batches several payments into one transaction therefore credits each of them once, even when two payers reuse a nonce, while one authorization presented twice (with or without from) is still credited once. If both presentations settle while the first one's credits are still being recorded (after a database error, or while another process holds the write lock), the one that settled first is still the one credited. Within one process this always holds; across processes it holds once the first one's settlement is marked settled in the database. The later one is answered 409 payment_already_redeemed and credited nothing. If the database refused that mark too (as it does while the lock is held), a sibling process cannot see the first settlement, and whichever presentation is recorded first is credited. npm run settlements -- resolve does not apply this rule (see Reconciling x402 settlements). If no usable payer is known for an authorization sent without from (the facilitator reports none, or reports it in any other form), it is recorded under its nonce alone (x402_<tx>:<nonce>). That record blocks the payer-qualified one, and any payer-qualified record of the same transaction and nonce blocks it, whichever of the two presentations arrives first. Records written by earlier releases (x402_<tx>) still block. So does a record whose payer part is not a 20-byte address, such as a hashed CAIP-10 account id that an earlier release wrote (or that a server or npm run settlements writes when it finishes a marker such a release left): it counts as a record without a known payer, so it blocks the same authorization presented with from, and with it another payer's authorization that reuses its nonce in the same batched transaction, since it may be the same authorization.
• The monitor's entitlement pass (pausing, re-planning, resuming and retiring self-serve watches) takes the write lock only when an account's watches need a change. When another process holds that lock, the rest of the pass waits for the next round instead of failing the round.

Publishing to MCP marketplaces

1. npm: bump version in package.json and both version fields in server.json (the top-level version and packages[0].version), then npm publish --access public. prepublishOnly rebuilds and runs the tests. The mcpName field ties the npm package to the registry entry below.
2. Official MCP Registry (registry.modelcontextprotocol.io): edit server.json so name uses your GitHub namespace (io.github.<your-user>/stack-and-dns-mcp), add your hosted remotes[].url, then mcp-publisher login github, mcp-publisher validate and mcp-publisher publish. The registry verifies ownership of the npm package through mcpName (keep it equal to name), and its schema limits description to 100 characters.
3. Cursor directory / Claude connectors / Glama / PulseMCP / mcp.so: submit the GitHub URL and the hosted /mcp endpoint; each directory reads server.json or the README config blocks.
4. Docker MCP Catalog: submit the image built from the Dockerfile with the --http entrypoint.

Development and testing

npm run dev          # stdio with tsx (no build)npm run dev:http     # HTTP on :3333npm run typechecknpm test             # node:test suites, offline (analyzers, scoring, auth, payments and ledger, bulk jobs, monitor, crawler, scan store)npm run format       # Prettier (.prettierrc.json); npm run format:check only reportsnpm run smoke -- example.com
# real MCP client round trip (initialize, tools/list, tools/call) over each transport, after npm run buildnode scripts/e2e-client.mjs stdio example.comSDM_KEY=sdm_live_... node scripts/e2e-client.mjs http example.com   # against localhost:3333 (override with SDM_URL)

npm test, and therefore npm publish (prepublishOnly), needs Node.js 22.13+: the suites use node:sqlite, and on Windows, where npm runs scripts through cmd.exe, only Node's own --test expands test/*.test.ts. The suite never connects off the machine: DNS is scripted, servers listen on 127.0.0.1, and the HTTPS end-to-end tests make their self-signed loopback certificate in process (test/fixtures/loopback-cert.ts), so they need no openssl binary and none are skipped.

Project layout:

src/  index.ts                 entrypoint, transport selection, storage and ledger attach, process guards  config.ts                Zod-validated env/CLI configuration and plan limits  server.ts                McpServer factory and instructions  schemas.ts               tool input/output Zod contracts  services.ts              storage-backed services and background workers (bulk, monitor, crawler, maintenance)  logger.ts                JSON-lines stderr logger  tools/                   inspect_tech_stack, verify_email_infrastructure, comprehensive_audit,                           domain_history, bulk checks, monitoring, reverse lookup  signatures/              signatures.json (450+), schema, compile + match engine  middleware/              auth (API keys, session tokens), payments (Stripe, x402), rate limiter,                           per-address throttle, usage ledger, tool-call meter  transports/              stdio, Express (Streamable HTTP + SSE) and the /v1/jobs REST API (jobsApi.ts)  store/                   SQLite schema and migrations (v1-v10), stored scans, bulk jobs, monitor,                           crawler, reverse index, credit ledger and x402 settlement markers,                           self-serve keys, worker lease  billing/                 self-serve key signup and the Stripe API client  cli/generate-key.ts      API key generator (npm run keys:generate)  cli/settlements.ts       operator reconciliation of x402 settlements (npm run settlements)  lib/                     domain parsing, DNS client with fallback, HTTP fetcher, TLS inspector,                           SPF/DKIM/DMARC/MX/MTA-STS/TLS-RPT/BIMI analyzers, security headers,                           WAF detection, scoring, SSRF guards, robots.txt, domain lists, webhooks  types/                   ambient type declarations  data/disposable-domains.jsontest/                      node:test suites (test/fixtures/ holds shared helpers)scripts/smoke.ts           run tools directlyscripts/e2e-client.mjs     real MCP client round trip over stdio or HTTP

Troubleshooting

MIT License. Copyright (c) 2026 Mewnah.