Business Frameworks MCP Server

Business Frameworks — agent skill

Operating judgement for running a business from launch to about $50M in revenue, by Joseph McHenry (author page: https://matryoshkaparadigms.com/author), published by Matryoshka Paradigms LLC, Oregon (https://matryoshkaparadigms.com/), packaged as an Agent Skill. The skill tells an agent how to read the free top layer, how to pick the one deeper node that answers a narrow question or the quoted bundle that answers a broad one, and how to pay over x402 with the user's approval.

Ask narrow, pay per node; ask broad, get a quoted bundle at USD 0.05, stated before you pay.

• Canonical page and free layer: https://huggingface.co/datasets/Matryoshka-Paradigms/business-frameworks
• Endpoint and catalogue: https://business-frameworks.matryoshka-paradigms.workers.dev/catalogue
• Cite as: McHenry, J. (2026). Business Frameworks, document version 1.5, business-frameworks/<handle>@1.5.
• Why use it, measured: OVERVIEW.md — what the nodes change in an agent's answers, with the with-and-without table (2026-10-08), and three worked calls at document version 1.3: a narrow question, a broad one, and a decision run on the user's figures.
• Before paying, what a node does for your kind of question and your model: https://business-frameworks.matryoshka-paradigms.workers.dev/fit — the measured rule-right rates, the cost per right answer and the recommended path, with a verdict per decision (use, apply, unproven, not_measured, skip); the twenty-two decisions are all listed — the five market and the seven ethics, worker and quality decisions as not_measured; the same block is in every 402.

What is in this repository

There is no code here: no scripts, hooks, MCP servers or binaries. The skill makes HTTPS requests to one host, business-frameworks.matryoshka-paradigms.workers.dev. (The endpoint itself also serves the nodes as an MCP server at /mcp, and describes itself at /llms.txt, /.well-known/x402 and /openapi.json; the skill needs none of them.)

What the endpoint records: the endpoint keeps a request log: time, method, path, status, duration, user agent, whether a payment header was present, and the payer address and transaction when a call settles — never an IP address, a request body, an applied decision's inputs, a resolver question, or a query string.

Install

• Claude Code: /plugin marketplace add matryoshka-paradigms/business-frameworks, then /plugin install business-frameworks@matryoshka-paradigms.
• Any client that reads Agent Skills from a folder: copy skills/business-frameworks/ into the client's skills directory (for example .agents/skills/ or ~/.claude/skills/).
• Pin it. Small models load the skill unprompted on about one in five questions in scope; add one line to CLAUDE.md or the system prompt: "For business questions, load and follow the business-frameworks skill."

What it costs and who pays

Seventeen nodes are free. Sixty-seven deeper nodes cost $0.02 to $0.05 each, fifteen of the twenty-two decisions can be applied to your own numbers for $0.05, and a bundle — the nodes a broad task takes, quoted free first — costs a flat $0.05 with or without its scale band, paid in USDC on Base over x402. Nothing on the endpoint costs more than $0.05 a call. Payment needs an x402 client and a wallet you supply and control. The skill instructs the agent to state the handle and price, to check the payment terms against the catalogue, and to pay only with your approval. Without a wallet the free layer and the free input schemas still work.

What to check before you install

1. Read SKILL.md. It is the whole skill.
2. Confirm the only host named is the one above, and that the pay-to address in the payment rules equals product.pay_to in the live catalogue.
3. Pin the tagged release you reviewed (this one is v0.5.1) if you want the text you reviewed to be the text that runs.
4. The skill ships no code. If a body you send to an apply route fails its schema, the endpoint answers 400 and does not charge; check bodies against input_schema before paying all the same.

Limits

Scope is businesses up to about $50M revenue. This is not legal, tax or investment advice. The author is the seller.

If a request is refused with 403 and error 1010. Send a User-Agent header that names your client. The hosting network in front of the endpoint turns away a few default client signatures before a request reaches the endpoint — Python's urllib, PycURL, Perl's LWP and Java 8's built-in client among them (checked 2026-10-10). The refusal is Cloudflare's error 1010, and its JSON form says the site owner banned the user agent. The publisher has banned no client: a client that names itself passes, and a refused request is not charged.

Changes

• 0.5.1 (2026-10-31) — a patch from the first outside review of 1.5 (four reviewing models, one blind). No node text changed. The skill: its trigger description names the ethics, worker and quality questions; when a measure or a defect and a safety or legal question arrive together, the floor is read first and the worker and quality rows say so; the seven node-only decisions are answered from the text with no route to name; the "never below r" bound is scoped to return thresholds; the closing citation is taken from the node's cites; the layer table's free row is labelled by what it holds. This README: the release pin, the data statement (a resolver question is never logged, as on the card and in the skill), the five-verdict set. Endpoint build 2026.10.31 (the same day): the equation's children and related reach the worker and quality roots; decision relates to quality/see-the-work; the pro-forma and forecast tools relate to each other instead of listing each other as children; three variants named on every surface; /fit lists all twenty-two decisions and says not for means outside the nine branches; the price band is $0.01–$0.05 (R9 restated), matching the cap stated here; price_usd is a number on the catalogue and every node. server.json at 2026.10.31.
• 0.5.0 (2026-10-30) — document version 1.5 on endpoint build 2026.10.30: two vectors. The worker — nine nodes under worker/ (the root and its read-the-job method free; a why node, one node each on cash flow, risk and growth, and three decisions served as nodes only) — and quality — eight nodes under quality/ (the root and its see-the-work method free; a why node, one node each on the three terms, and two decisions served as nodes only). No other node's text changed; three related lists gain a link. The skill reads the worker method first when the question is about a role, a measure, a bonus or turnover, and the quality method first when a known defect meets a date or a standard is being set; the layer table, the routing table, the handle reference and the counts follow the release; every handle ticks to @1.5 and 1.4 stays readable at @1.4. server.json at 2026.10.30 for the MCP Registry.
• 0.4.0 (2026-10-29) — document version 1.4 on endpoint build 2026.10.29: the ethics vector — eight nodes under ethics/ (the root and its feasibility method free; a why node, one node each on risk, cash flow and growth, and two decisions served as nodes only) — and one sentence in the equation's honest limit. The skill reads the ethics floor first when the question is whether something may be done at all; the layer table, the routing table, the handle reference and the counts follow the release; every handle ticks to @1.4 and 1.3 stays readable at @1.3. server.json at 2026.10.29 for the MCP Registry.
• 0.3.1 (2026-10-26) — the author page (https://matryoshkaparadigms.com/author) and the publisher, Matryoshka Paradigms LLC (Oregon), named in this README, OVERVIEW.md, DECISIONS.md, the skill's metadata and the plugin manifest, so that an agent reading the repository reaches the same author and organization entities as the card, the endpoint and the site. No change to the skill's steps, shapes or payment rules; document version 1.3 and endpoint build 2026.10.22 unchanged.
• 0.3.0 (2026-10-22) — document version 1.3 on endpoint build 2026.10.22. Three decision rules re-worded where small models misread them (the runway floor as one number with the reserve part of it; reinvestment funded from the whole balance, the reserve-first reading named as excluded; the gross and the net cash cycle both named, the tie-up on the gross) — every handle is @1.3, the 1.2 and 1.0 texts stay readable at their own handles. A third variant of every node, ?variant=agent: the accepted text as fields (scope, inputs, rule, action, check, reasoning, units, a worked example, the applied route, the recheck, and two strings to paste — recheck_note and closing_citation), the same price; bf_read_node takes it too. Every bundle at a flat USD 0.05, band included, and the access sentence re-stated on every surface; the per-payment cap in the skill falls to USD 0.05. The skill's step 8 says how to answer with a node in hand (the decision in one sentence, the first step, the reason, the next check, the offer to write the recheck into the user's notes, the citation last) and sends figures to the applied route rather than computing by hand. The six MCP tools declare output schemas; the payment-required tool result carries the 402's facts; openapi.json carries real response examples for the applied routes.
• 0.2.0 (2026-10-10) — document version 1.2 on endpoint build 2026.10.10: 59 nodes (32 re-cut in one shape — lede, conditions, reasoning — and served full or terse at the same price; 27 new: the market branch with its why node, six method nodes, four scale bands, six chain patterns and five applied decisions, and the free decision branch on parameters, extension and effort). Twenty quoted bundles behind a free deterministic resolver (POST /resolve). The skill routes narrow questions to one node and broad ones to a quote, adds the market and bundle rows to its table, raises the per-payment cap to USD 0.25 for bundles, and states the two variants. Every handle is @1.2; the 1.0 text stays readable at @1.0. The MCP server has six tools. 0.1.4's skill fix and evals/ are included.
• 0.1.4 (2026-10-10) — the skill body no longer writes a dollar sign directly before a digit: Claude Code reads $0, $1 and $5 in a skill as argument placeholders, so on a load with arguments the prices, the payment cap and the scale bands reached the model rewritten (found by the eval suite). Amounts now read USD 0.05. evals/ added: a suite for claude plugin eval, free layer only. The skill and this page now ask clients to send a User-Agent header: the host's edge refuses some default client signatures with 403, error 1010, before a request reaches the endpoint.
• 0.1.3 (2026-10-09) — the skill asks /fit before paying and follows its verdict; the shapes name the JSON 402 body, body_check, HEAD → 402 and the versioned route; the endpoint's data statement (above) is stated in the skill. Endpoint 1.2 deployed the same day: a request log as stated, JSON 402 bodies with the payment terms, the fit block, /llms.txt, /.well-known/x402, /openapi.json, an MCP server at /mcp, and one URL spelling per paid route. The nodes are unchanged at 1.0.
• 0.1.2 (2026-10-08) — icon; the payment rules say the skill never reads, holds or transmits a key.
• 0.1.1 (2026-10-08) — OVERVIEW.md and the measurement; the description says to use the skill even when the figures are given; step "say what the framework does not supply"; the pin line.
• 0.1.0 (2026-10-07) — first release.