Craft MCP

The editorial runtime for AI writing.

Teach Craft your voice once. It briefs the model from your best work, gates every draft against your rules, and learns from every verdict — one memory under Claude, ChatGPT, Cursor, and your CI. Yours to audit, export, and take anywhere.

01 / 05
  1. 1Prompt

    Use my voice to draft an opening for a short story about memory.

  2. 2Tool call · what Craft hands the AI
    craft_compose
    gathers your most relevant memories
    ranks them for fit to this story
    drafts a brief from your editorial memory
  3. 3Result · Your AI
    Memory is not a room you return to but a tide. It rearranges the furniture of the past each time it comes in, and leaves only what it chooses — a shoe, a sentence, the grey of a morning you thought you’d lost.

The runtime

Brief. Gate. Learn. Every draft runs the loop.

Not a writing app — a layer under whatever you already write in. Three moves run on every piece, and each one makes the next piece better.

01 · BRIEF

Drafts start from your evidence

For every task, Craft assembles a brief from your closest real passages — with the reason you approved them — plus your rules stated up front. The model writes from your best work, not from vibes.

02 · GATE

Checks are deterministic

No model grading a model. Every violation names its rule, shows the exact span, and carries a fix — style rules yield to intentional craft, policy rules never do. Same result, every run.

03 · LEARN

Your verdicts compound

Approve with a reason, reject with a complaint, or just hand back your edit — everything becomes memory. Repeat rejections surface as suggested rules; the next brief is already sharper.


Memory

Two kinds of memory. One you enforce, one you refine.

A prompt is just a wish. Craft separates the rules you can't break from the style you want to keep, and handles each one differently.

HARD RULES

Rules that never bend

Your house rules, compliance lines, and hard quality bars — banned words, length limits, required sections. Checked identically on every draft, with the exact break pointed out.

SOFT VOICE

A style that gets sharper

For every task, Craft builds a brief from your memory — only the most relevant examples steer the draft, so quality stays high and the voice keeps sharpening.

PER CHANNEL

Every register holds its own line

Your tweets, your essays, and your client's brand voice are different registers. Craft learns each channel separately, so punchy never bleeds into long-form — and drafts are checked against the right one.


Tech Commentary — No HypeMature

"Sharp, confident, intellectually curious. Concrete claims over vague gestures. Humor is dry, not forced."

12.9words / sentence
0.47sentence variety
73.5vocabulary range
demandingreadability
Hard memoryenforced on every draft
Never use "virtually limitless"from a rejection
Keep posts between 300 and 500 wordslength
Open with a concrete claim, never a definitionstructure
Soft memorysteers, never blocks
~Treat the reader as smart12 examples
~Dry humor, never forced8 examples
~Short sentences carry the turnsinferred
Suggested rule: ban "It works" — you rejected it twiceAdd rule ↗

See exactly what it learned. Change anything.

A measured snapshot of your voice — sentence length, variety, vocabulary — beside the rules you wrote and can read.

When you keep rejecting the same thing, Craft notices and suggests a rule. One click turns it into a check that's never broken again.


One memory, every tool

Teach it once. It holds everywhere.

Model memory is a vendor feature — opaque, and locked to one assistant. Craft memory is an asset you own: the profile you teach in Claude is the same one ChatGPT drafts from, Cursor edits with, and your CI enforces.

PORTABLE

Switch models freely

Connect Claude and ChatGPT with your account, Cursor with a key, CI with the API — one voice under all of them. A better model next year? Your memory comes along.

AUDITABLE

See everything. Remove anything.

Every example, rule, and correction is listed with its date and provenance — and a Remove button. The full teaching timeline shows how the voice was built, verdict by verdict.

YOURS

Export the whole voice

One click exports your style guide — fingerprint, constitution, exemplars — as plain markdown. Hand it to a colleague, keep it in your repo, or leave with it. No lock-in, by design.


Toolkit

Eight tools your AI calls on its own.

Each does one job — compose, check, fix, rank, teach. Pick one to see it in action; your AI runs them over MCP or a plain web API.

The brief

Start every draft from your best work.

Before your AI writes, Craft assembles a short brief for the task — your strongest examples, the rules that apply, and the habits to avoid.

  • Task-matched examples
  • Only the most relevant memories
  • Rules + habits to avoid
craft_compose
  1. 1Prompt

    Draft the opening for our Q3 launch post.

  2. 2Tool call · what Craft hands the AI
    craft_compose
    pulled 3 of 47 memories · ranked for this task
    Most AI demos work. Most AI products don’t. The gap isn’t technical — it’s operational.your saved example · note: “concrete opener, no hype”
    + 2 rules apply · 2 habits to avoid
  3. 3Result · Your AI

    Most launches lead with the feature. Ours leads with the workaround you’ve lived with for two years — and the afternoon it stops being your problem.


The playbook

Five minutes to running. Four pages to fluent.

The guide walks you from sign-up to your first checked draft — then goes deep, with copy-paste prompts for every move.


Start now

The layer no prompt or model can replace.

A bigger model, a sharper system prompt, a static skill — none of them remember how you write or hold the rules you can't break. Memory is the missing layer for consistent, high-quality prose from any language model.