Sendermap MCP Server

Sendermap

Type a domain, see every service that sends email as it, and the public evidence for each one.

Status: early release. The command-line tool works, and the website runs at sendermap.dns-audit.com. Needs Python 3.12 or later.

What it shows

For any domain, Sendermap builds a map of the services that send or receive its mail: mailbox providers, marketing and transactional platforms, support desks, security gateways and the services that receive its reports. Each one comes with:

• The evidence, record by record: the SPF include, MX host, DKIM key, bounce or tracking CNAME, verification token or report address that gave it away.
• How sure we are, in words, not a percentage (the terminal and the site show these as Receives your mail, Set up to send, Likely sends and Account only; the JSON keeps the tier names):
  • In use: mail for the domain is delivered there.
  • Configured: set up to send as the domain.
  • Likely: a signing key whose name points at the service, or the service's script, mention or subprocessor listing on the domain's website.
  • Account only: a trace of an account with nothing set up to send.

It looks beyond the main domain: brands often send from subdomains such as email., mg. or bounce., which a check of the main domain never sees. It also points out what needs attention: DKIM records that point at names that no longer exist, SPF records past the 10 lookup limit, revoked keys, and senders nobody can name. Raw IP ranges nobody can name are grouped by the network that announces them, so a dozen ranges from one provider read as one line.

Each DKIM key is shown with the tags that change what it does: t=y (testing, so verifiers treat mail signed with it as unsigned), t=s (strict), h= (hash algorithms), s= (service types) and k= (key type). A tag that weakens the key, such as testing mode or SHA-1 only, is raised as a finding. Each key also gets a fingerprint, fp and 16 hex characters, so the same key published at two selectors is easy to spot, and the map points it out when it is.

Looked for and not found

The map also lists what it looked for and did not find, with the DNS answers that show it. Today that covers DKIM keys. Whether "not there" can be trusted depends on how the domain's DNS answers for a name that does not exist, so at each name Sendermap first asks for a selector nobody publishes and reads the answer as one of four classes:

• Classic: the zone says "no such name". Answers can be trusted. If _domainkey itself does not exist, no selector exists under it. The map checks only the few names the domain's SPF and MX point to, rather than its full list, and says so. If _domainkey exists but every name asked is missing, the map says selectors exist but none of the names it tried is one.
• None by rcode: the zone gives an empty answer for every missing name (compact denial), so the answer says nothing about what exists. The map asks its full list and claims no absence.
• Wildcard: every name answers. The map ignores what the wildcard repeats and asks its full list.
• Unknown: the test lookup failed, so nothing is read from it, and the map lists it as not checked.

How it works

Everything comes from public sources: DNS, certificate transparency logs, the domain's own website, and public records of who owns each network. Each finding is shown with the record that supports it, so you can check it yourself.

• SPF is walked as a tree, crediting only links that actually authorize senders.
• DKIM keys are found by scanning the names services use, and credited by where they point. A plain name many hosts use, such as s1, credits its service (Klaviyo) only where the domain's SPF, MX or an NS delegation already shows that service.
• Raw IP ranges in SPF are named by matching them against the ranges each service publishes, then by who owns the network.
• Sending subdomains come from the names services ask for and from certificate logs. A subdomain delegated by NS to a service's nameservers, as Klaviyo and Salesforce Marketing Cloud set up, counts as that service's.
• The website shows the scripts and stylesheets it loads, the services it mentions, and the services its subprocessor page names. Every page fetch goes through one guard: https only, with the host's addresses checked first, so a domain cannot point a fetch at a private or internal address. A refused fetch is listed as not checked.
• Certificate transparency: Cert Spotter's log is read for the subdomain search, so on every map except one made with --no-subdomains. With --ct (off by default), crt.sh is read too, since it also holds expired certificates, and every logged name is resolved now and matched like any DNS record, so a branded tracking link such as link.example.com pointing at an email platform shows up. A logged name that no longer resolves but still points at a service is listed as a possible past vendor, not a current sender. crt.sh is often slow; it is tried three times, each log's answers are cached for a day, and a log that never answers is listed as not checked while the map finishes without it.

Run it from the command line. It installs from PyPI (Python 3.12 or later):

pip install sendermap
uvx sendermap example.com             # run without installing

The latest code, ahead of a release, installs from GitHub with pip install git+https://github.com/marmot7775/sendermap.

sendermap example.com
sendermap example.com --evidence     # every piece of evidence
sendermap example.com --markdown     # a summary to paste
sendermap example.com --json         # everything, for scripts
sendermap example.com --ct           # resolve every name in certificate logs, crt.sh too
sendermap example.com --claimed "Office 365, Mailchimp"   # hold the services you were told of against the map
sendermap example.com --claimed-file services.txt         # the same, one name per line
sendermap example.com --no-web --no-subdomains            # skip the website and the subdomain search
sendermap example.com --stop-early-dkim                   # faster DKIM probe that stops early
sendermap example.com --doh                               # look up DNS over HTTPS through Cloudflare
sendermap example.com --json --output map.json            # write the map to a file

For scripts and agents: the exit code is 0 when the map was built, 2 for bad input and 3 when the domain's own records could not be read (no map). Progress goes to stderr, and only to a terminal, so stdout holds the map alone. sendermap --help explains every flag.

The website

sendermap.dns-audit.com builds the same map on a web page, with progress shown while it runs. The form has a second box, Services you use (optional): name them, comma separated, and the map ends with Your list against the map, the same check as --claimed. Each map has a link to share, ?d=DOMAIN, with &claimed= added when names were given, plus buttons to copy the Markdown or download the map.

The site's API gives the map as JSON (/api/map?d=DOMAIN), Markdown (/api/map.md), the HTML fragment the page shows (/api/map.html) and a server-sent event stream (/api/map/stream). Each takes &claimed=. Requests are rate limited per visitor and per target domain. /llms.txt lists the API, the schema, the MCP endpoint and the skill for agents.

The site's request log records the path, status and timing of each request, without the query string or the visitor's address, so the domain looked up is not in it. A built map is held in memory for ten minutes so a shared link opens at once. Claimed names are used for that response only and are not cached with the map.

MCP

With the mcp extra, sendermap mcp serves two tools over MCP on stdio: sendermap_map, whose structured content is the map as sendermap/1 JSON (summary or full), and sendermap_vendor, one vendor's patterns, proofs and documented setup. The schema and the vendor list are resources.

{
  "mcpServers": {
    "sendermap": { "command": "uvx", "args": ["--from", "sendermap[mcp]", "sendermap", "mcp"] }
  }
}

The website serves the same tools at /mcp over streamable HTTP. For agents without MCP, skills/sendermap/SKILL.md is an Agent Skill that runs the command line and reads the JSON.

Honest limits

• No volume. Public records show which services are set up to send, not how much mail each one sends, or whether it sent any this month.
• No proof a key is in use. A DKIM key in DNS shows it was published, not that anything signs with it. Only real mail shows that, and Sendermap reads no mail.
• No sight of vendors sending from their own domain. A service that mails your customers from its own domain, such as a receipt from a payment platform, leaves nothing in your DNS.
• Selector discovery is never complete. DNS cannot list DKIM selectors. Sendermap asks the names services are known to use and some common ones, and says how many it asked. A key under any other name stays unseen.
• The selector probe can stop early, when asked to. By default every selector is asked. With --stop-early-dkim it asks first the names of every service SPF, MX and CNAMEs already name, the generic names, and the names of services that need no SPF include. Once every service SPF, MX and CNAMEs name has a key, the rest is not asked, and the map lists those names as not checked, never as not found.
• Absence is read from how the zone answers, and said as such. "Not there" is claimed only where the zone's denial class makes the answer trustworthy, and every such line carries the answers it rests on.

The JSON contract

sendermap DOMAIN --json and the site's API give one map as JSON, schema sendermap/1, described by sendermap/data/schema/sendermap-1.json (JSON Schema 2020-12; sendermap.schema() in Python, /schema/sendermap-1.json on the site). sendermap/1 grows by addition only: new fields and new values (a new kind of evidence, a new gap kind) may appear, and nothing is removed, renamed or changed in meaning. Ignore fields you do not know and tolerate values you do not know. A breaking change would be sendermap/2.

The vendor data

All vendor knowledge lives in one file, sendermap/data/vendors.json, one block per service, so adding or correcting a vendor is a small pull request. The first version was imported from the vendor tables in dns-audit (MIT).

Contributing

Each vendor gets one block in sendermap/data/vendors.json, and everything known about it goes there. Every pattern in a block (an SPF include, an MX host, a CNAME target, a report address, a verification token, a network, a website host, a selector name, a subprocessor name) carries one proof of its own: a live domain that publishes it, the record name or page as found, and the date checked (a numbered series of selectors such as s1, s2 needs one for the series). A pattern without a proof does not get merged, and Sendermap refuses to load one. A block may also list aliases, the other names customers call the service by ("Office 365" for Microsoft 365). An alias needs no proof, since it matches only what a person types, never a record.

Patterns not yet proven wait in sendermap/data/vendors_pending.json, each block with a note of what was checked; Sendermap never reads that file, so a record they would match shows as not recognised. A vendor moves from pending to vendors.json with one proof per pattern, the same rule as any new block. tools/hunt_proofs.py looks for those proofs on the customer domains in data/hunt_domains.txt, with the same checks as the weekly recheck.

Which vendor to add next comes from the benchmark: tools/benchmark.py maps the same 300 domains each release and ranks what it could not name by how many domains publish it. On 0.1.7, 83 such groups showed at two domains or more.

Every proof is rechecked against live DNS and the web each week and on any pull request that touches vendors.json. A proof that no longer holds opens an issue, so a stale pattern is caught and fixed.

A tag v* builds the release files and, while the repository variable PYPI_READY is true (Settings, Secrets and variables, Actions, Variables), publishes them to PyPI by trusted publishing and makes the GitHub release with the tag's notes from CHANGELOG.md. The same run lists the version in the MCP registry from server.json and upgrades sendermap.dns-audit.com to it. deploy/README.md covers the site's server files.

Related

• dns-audit.com: a full email security audit of a domain.

License

MIT