Skip to content
How we build/The judgment calls

The rules we keep when nobody is checking.

On October 11, 2026 we shipped 14 free connectors for Claude in one day. The speed was the easy part. These are the calls we made along the way, each with the receipt.

Read the callsHow we set up Claude

Counted from the repos and the live servers on October 11, 2026

A mustard seed in a cap holds up a ring of brass keys in front of a vault door
You hold the keys. We keep the rules.
  • 14Claude connectors shipped in one day
  • 69tools across them
  • 513unit tests, all passing
  • 171live checks against production, all passing
  • 14listed in the official MCP Registry

Nine of our tools are listed in Anthropic’s Claude directory. See them all

Claude is a product of Anthropic. Modern Mustard Seed is an independent studio and is not affiliated with or endorsed by Anthropic.

The same rules, at scale: four of these engines measured 2,000 small business websites for the Small Business AI Readiness Index, aggregates only, no business named.

9 calls/Each with its evidence

Fast is easy. Right is the job.

Every call below cost us something: a popular data set, a convenient example, a sales line, a shortcut. Links open the public source. Commit references point to the change in our own repositories.

  1. 01Church Year

    We shipped the readings we were allowed to ship, not the popular ones.

    Most churches follow the Revised Common Lectionary, so that was the obvious table to serve. We read its permissions page first. Any organization other than a congregation needs written permission from Augsburg Fortress to deliver it electronically. So Church Year serves the 1979 Book of Common Prayer lectionary instead: public domain, the parent of the RCL, the same three-year cycle and Sunday names.

    WhyA free tool is still a publication. We do not get to borrow someone else’s copyright because the tool costs nothing.

    Evidence
  2. 02Call Desk

    Every legal rule carries its statute, word for word.

    Call Desk answers whether a business can place a call right now, and whether it can record it. Every rule it applies comes back with the citation, the verbatim text of the law, the official source link and the date we read it. Ask about a Pennsylvania number for an evening after October 18, 2026 and it applies Act 47 on the day it takes effect: 9 a.m. to 7 p.m., Monday through Saturday, never Sunday.

    WhyA business owner should never have to trust us. The answer shows its work, so they can check it in one click.

    Evidence
  3. 03Call Desk

    If we could not read it on an official source, we do not apply it.

    Some state rules show up only in news reports, old bills or secondary summaries. We list those in the tool’s answer, with the reason each was left out, and never apply them. Today that is nine calling-hours rules and four AI disclosure claims. The gap stays visible instead of being guessed.

    WhyA wrong “yes, you can call” costs a small business real money. An honest “we could not confirm this one” costs nothing.

    Evidence
  4. 04Cross + Covenant Bible

    Scripture only, in a translation we are allowed to serve.

    Our daily devotionals quote writers like Lewis, Tozer and Bonhoeffer, and hymn writers too. That works on our own site. Served to anyone through Claude, those quotes are not ours to hand out. We made 46 connector-only edits that remove each quote and rewrite the sentence around it, and caught one verse in a modern translation’s wording and set it back to the public-domain World English Bible. A test now fails if a served devotional names a non-scripture author.

    WhyFaith content deserves more care, not less. The rule is in a test, so it holds after we stop looking.

    Evidence
  5. 05Examples

    No real phone numbers, not even ours.

    An early draft of the Call Desk examples used a real personal cell number. We replaced every example number in the docs, listing prompts, evals and tests with the 555-0100 to 555-0199 range, which is reserved for fiction. The tool flags those numbers as fictional and still answers for their area code.

    WhyAn example prompt gets copied thousands of times. A real number in it is someone’s phone ringing.

    Evidence
  6. 06Security

    Our servers refuse to be pointed inward.

    Several connectors fetch a website someone names. That is the classic way to trick a server into reading its own network. Every fetch goes through one guard that refuses private, loopback, link-local and reserved addresses, alternate spellings of them, odd ports, and redirects that land inside a private network. Thirteen tests cover it, and we try it on production: ask Site Check to scan the cloud metadata address and it refuses, by name.

    WhyA free tool open to anyone in Claude is open to everyone. The guard is the cost of saying yes to that.

    Evidence
  7. 07Directory policy

    The tool gives the answer, never a pitch.

    Anthropic’s directory policy rejects connectors that work as a promotional vehicle. So every answer is the data and one attribution line: no price, no offer, no link to buy. We kept two of our own engines out entirely because their output led to a sale. Example prompts name national brands, never a small local business that did not ask to be a demo.

    WhyA connector that sells is a worse connector. The help is the point.

    Evidence
  8. 08Privacy

    We record what was scanned, never who asked.

    The connectors have no login, and we keep it that way: no prompts logged, no identity guessed, no fingerprinting. When we add usage counts, they will record the public website that was checked, never the person in Claude who checked it, and the privacy policy changes before anything is recorded. A person becomes a contact only by writing to us themselves.

    WhyPeople will paste their business into these tools. That trust is the asset, so we protect it in the design, not in a promise.

    Evidence
  9. 09What we say about us

    We state our size and our standing exactly as they are.

    We are a small studio. We do not count software agents as employees, we do not call ourselves certified for an exam we have not sat, and nothing we publish says Anthropic endorses or partners with us. Our Claude lessons carry that line in every caption.

    WhyEverything else on this page is only worth reading if this one holds.

    Evidence
The defaults/On every connector

Four things we do without being asked.

  • No model inside the tool

    The connectors run on rules, public data and arithmetic. Same question, same answer, and nothing to bill per call.

  • Limits per tool, never per person

    All Claude traffic arrives from one network range, so we cap each tool and each target, with a daily ceiling.

  • Read-only by default

    Every tool is marked read-only except the ones that change your own saved data, and those say so.

  • Tested before it merges

    Typecheck, lint, build and unit tests on every change, then real prompts run against the live server.

Your business, next

Want Claude set up with the same care?

We set Claude up for businesses and run it with them: skills for the work you repeat, rules it follows, guardrails it cannot cross, and your tools connected. Managed for you, end to end, and every account stays in your name.

Begin a conversationSee the Claude setup

Claude is a product of Anthropic. Modern Mustard Seed is an independent studio and is not affiliated with or endorsed by Anthropic.

[ Related ]

Keep reading