Hammer Worlds for Claude, Codex, and Agent Skills

A hammer world is a corpus that publishes what it cannot answer. It arrives as an MCP server, it hands out figures with receipts attached, it checks the claims you write against those receipts, and where a question is outside what its sources can settle it refuses and says why. The refusals are not a limitation that came with it. They are the product.

This repository distributes one plugin across Claude Code, Claude Cowork, Codex, and clients that implement the Agent Skills format. Every native plugin install mounts the hosted coverage-intelligence world and carries one skill about the discipline rather than any particular corpus.

New here? The release note is the place to start, and it is written for people who read the literature closely and then have to defend what they write: Hello, Hammer Worlds. It covers what the corpus holds, why a refusal is an answer, and what this will not do.

What is in this release

The coverage-intelligence world serves what commercial payers published they pay, code by code, alongside the Medicare fee schedules, the NCCI billing-pair edits and the Medicare Coverage Database articles that qualify a rate. Every figure arrives with the number of observations behind it and two figures saying how thin that evidence is, and a cell below the reporting floor is refused rather than averaged into something that looks stable.

Five tools handle the case where your input is a range rather than a number: run a world's own estimator over distributions you declare, rank which of your assumptions actually moves the answer, check a number you already believe against what the model produced, ask what would have to change, and build a prior from a small declared network. Everything they serve is marked as resting on assumptions, even where the underlying figure was measured, because arithmetic over your own declarations does not become an observation.

The skill in this repository is about the discipline rather than the corpus: how to route a question through whatever worlds a client has mounted, why two worlds' answers never compose into a third claim, how to read a refusal, and the claim report every answer built on a world has to close with.

If you installed before 25 August 2026, reinstall

Your client is pinned to a version it can no longer update past, and it will not tell you. It will keep showing whatever it installed, with no error.

On 25 August the plugin moved from a hammer-worlds/ subdirectory to the root of this repository, so the marketplace entry's source changed from ./hammer-worlds to ./. An install made before that points at a path that no longer exists. Every release since, 0.2.0 through 0.4.0, has been published somewhere those installs do not look.

How to tell. Open the plugin's details. If the version reads 0.1.0, or Last updated says 24 August or earlier, you are on the orphaned install. The current version is 0.4.0.

Nothing is wrong with what you have; it simply stopped receiving updates, which in this case means it is missing the five tools that take a range instead of a point.

The fix is to remove and re-add. An update will not do it, because the update follows the same dead path.

Claude Cowork

Customize > Plugins, open Hammer Worlds, use the ⋮ menu to Uninstall. Then remove the hammer-plugin marketplace if it is listed separately, add hmmrlabs/hammer-plugin again, and install Hammer Worlds. Confirm the version now reads 0.4.0.

Claude Code

/plugin uninstall hammer-worlds@hammer
/plugin marketplace remove hammer
/plugin marketplace add hmmrlabs/hammer-plugin
/plugin install hammer-worlds@hammer

The marketplace removal is the step that matters. Re-adding without it can reuse the cached entry that carries the old path.

Codex

codex plugin remove hammer-worlds@hammer
codex plugin marketplace remove hammer
codex plugin marketplace add hmmrlabs/hammer-plugin
codex plugin add hammer-worlds@hammer

Agent Skills clients

Re-run the install command below; it fetches the skill by path from the current default branch rather than from a cached marketplace entry, so it was not affected.

After reinstalling, reconnect the MCP connection. A client that connected before a world deploy holds the old tool list, which is a separate cache from the plugin one.

Install it

Claude Cowork

Open Cowork > Customize > Plugins. In Personal plugins, choose +, then Add marketplace. Add hmmrlabs/hammer-plugin, install Hammer Worlds, and complete the connector sign-in when prompted.

Claude Code

/plugin marketplace add hmmrlabs/hammer-plugin
/plugin install hammer-worlds@hammer

Choose user scope when prompted, so the world is reachable from every directory rather than from one repository. The skill explains at length why that is not a matter of taste; the short version is that a silently absent tool is worse than a broken one, because the answer still arrives and still looks like an answer.

Then run /mcp and authenticate. A browser opens. There is no key to paste, no token in this repository and nothing to put in your settings.

Codex

codex plugin marketplace add hmmrlabs/hammer-plugin
codex plugin add hammer-worlds@hammer

Approve the OAuth prompt during installation. If authentication was deferred, open the plugin or MCP connection in Codex and authenticate there. Codex reads the same root skill and the same .mcp.json as Claude.

Other Agent Skills clients

npx skills add hmmrlabs/hammer-plugin --skill hammer-worlds -a <profile> --yes

Replace <profile> with the client profile supported by the Skills CLI. This installs the skill. A skill-only client may not install MCP dependencies from agents/openai.yaml; when it does not, add the remote HTTP endpoint https://worlds.hammer.ai/mcp through that client's MCP settings and complete OAuth there.

Three layers remain separate. MCP provides the world and its tools. The skill provides the procedure for using those tools without overstating them. Plugin manifests distribute both together where a client supports plugins.

What it costs

Nothing to run. The world is hosted at https://worlds.hammer.ai/mcp and it is the operator's own service, so there is no build, no download and no corpus on your disk. What a given caller is entitled to ask it is a separate question from being able to reach it: signing in says who you are, and the world decides per compartment what it will serve you. Where you are not entitled to something it says so and names the way to ask, through its request_access and redeem_code tools. A refusal for want of entitlement is a refusal with a remedy, not an outage.

What it refuses

This is the part worth reading before you install anything. Three of the questions people most often bring to a rates corpus, and why this world will not answer them:

• It cannot tell you a denial rate, or a probability that a claim gets paid. Not by any method, including arithmetic over counts it will happily serve you. Machine-readable files hold negotiated rates, not adjudications. A rate being present does not record a payment and a rate being absent does not record a denial, so there is no denominator anywhere in this corpus for that fraction.
• It cannot tell you whether a payer covers a code. A published negotiated rate says what a payer pays when it pays, not whether it will. "Payer X covers code Y" is not a sentence these sources can settle in either direction, and the absence of a rate row is not evidence of non-coverage: payers omit, fragment and misfile their disclosures market-wide. The supportable sentence is that no rate was recorded in this month's disclosure.
• It cannot turn relative value units into dollars. A code priced nationally on the Medicare Physician Fee Schedule carries relative value units, and converting those to money needs a conversion factor and a locality adjustment this world does not hold. So it returns the units and no ratio, rather than a dollar figure that would be a guess wearing a decimal point.

It also declines, among others, to average across payers without weights you declare, to predict or rank an individual provider, to make any clinical claim about any device or test, to state what was in force after its own vintage, and to add a cost that arrives later to a cost that never arrives.

That is a summary, and the world publishes the real list itself. Its full cannot_answer list arrives in the MCP initialize handshake as the server's instructions, and every tool carries its own refusals array, so a client can see what the world declines before spending a call. Where this README and the handshake disagree, the handshake is right: it comes from the build that is answering you and this file does not.

Why this plugin is so small

One skill and one URL. That is deliberate, and it is the plugin's best argument.

A world ships its own instructions and serves them from its own deployed build. Each world carries a skills/ compartment inside its package, and the server exposes it over MCP: resources/list returns one resource per skill at hammer://<package>/skills/<skill>, and resources/read returns that SKILL.md out of the running image. The coverage world's routing skill, hammer-coverage-navigate, comes to you that way.

So the routing instructions for a world and the tools that world answers with are the same revision by construction. They cannot describe a version that no longer exists, because there is only one of each and they are deployed together. If you have ever installed a plugin whose documentation described a release from six months ago, that is the failure this arrangement removes rather than manages.

What is left over for this repository is the part that belongs to no world: how to route a question through whatever worlds a client has mounted, why two worlds' answers never compose into a third claim, how to read a refusal, and the claim report that any answer built on a world has to close with. That is the hammer-worlds skill, and it is the only skill here.

What this plugin sends where

The dialog that adds a marketplace warns that Anthropic cannot verify third-party plugins. That warning is correct and this section is what it is owed. Everything in it can be checked against the files in this repository before you install anything, which is the only kind of assurance worth offering.

• Read it first; it is small enough to read. One skill, one YAML interface declaration and five JSON manifests. Nothing here executes.

• No hooks, no commands, no agents, no binaries, no corpus. Nothing in this plugin runs on your machine on a trigger you did not press, and nothing in it can read a file, because there is no code in it that could.

• One network destination, declared in one file. .mcp.json names https://worlds.hammer.ai/mcp and nothing else. What crosses the wire is the tool call your client made and the arguments you put in it.

• The service is the operator's own, running on Google Cloud. Its replies carry a server: Google Frontend header, and its authorisation is OAuth against an authorisation server named in its published protected-resource metadata, which you can fetch yourself before installing anything:

  curl https://worlds.hammer.ai/.well-known/oauth-protected-resource
  

  Measured 2026-08-24: an unauthenticated POST to the MCP endpoint answers 401 with a www-authenticate: Bearer header naming that document. That is the correct answer to an anonymous caller.

• The operator keeps a ledger. Every answer the world hands out is a receipt, and every claim submitted against one is logged with its verdict, on the operator's side. That exists so that answers taken and never checked are countable rather than assumed. If that is not a trade you want to make, do not install this.

What is deliberately not here

No corpus travels in this repository, and none ever will. The data the coverage world reads sits behind licence conditions that permit acquisition and not redistribution: some of it is cleared for non-commercial use with a source linkback, and some of it is cleared to hold and not to pass on. A public repository shipping any of it would breach that. The plugin ships a pointer and the discipline for reading what comes back, and the world serves the data under its own terms or declines to.

The noise world is not here either, and it is a world an evaluator would enjoy: it measures how much of a set of judgements is the case and how much is whoever picked it up, and it refuses more than it answers. It is absent for a plain reason rather than a policy one. Measured 2026-08-24, the hosted service answers on one endpoint and paths under it named for other worlds return 404, so there is nothing for a second entry in .mcp.json to point at. Shipping the noise skills without a noise server would hand an agent a procedure for reading a payload it has no way to obtain, and the measured failure that procedure exists to prevent is precisely an agent doing the arithmetic by hand and presenting the result. When the world is hosted, it becomes one more entry in the same file.

Licence

Apache License 2.0, in LICENSE, with NOTICE stating what it covers.

The skill here is prose rather than code, and a content licence such as CC BY 4.0 would fit prose more exactly. Apache-2.0 is used anyway, for one reason that outweighs the fit: this text moves between this repository and the private one that holds the hammer runtime, which is Apache-2.0, and the same text under two licences is a worse problem than a licence that is slightly the wrong shape for its subject. Attribution is required either way, and the patent grant and the explicit contribution terms are useful for the JSON manifests that sit beside the prose.

The licence covers what is in this repository. It says nothing about the corpora the hosted world reads, which are not distributed here and travel under their own terms.