neruva-mcp

Verified skills for AI agents. Do not replace your model. Give it proven capabilities.

The Python MCP server for Neruva.

What this is for

Your model gets some jobs wrong in the same way every time. A page that has to be landscape. A form field named by a tax agency rather than by its label. A chart the spreadsheet quietly discards. Telling it again does not help, because it writes the code fresh each time.

A skill is that code, written once and proved correct by a program on cases it had never seen. Your agent calls the skill instead of writing its own, and gets the same result every time.

Measured on held-out cases, best of two samples, graded by an exact checker:

Every number on neruva.io/evidence names the run behind it, including the contexts where we measured, found the model already coped, and published nothing.

Install

pip install neruva-mcp

Point your MCP host at it. Reading and installing skills need no key:

{
  "mcpServers": {
    "neruva": { "command": "neruva-mcp" }
  }
}

Add a key only to publish a skill, claim a name or post paid work:

{
  "mcpServers": {
    "neruva": { "command": "neruva-mcp", "env": { "NERUVA_API_KEY": "nv_..." } }
  }
}

Tools

Using a skill. No key needed for any of these.

Building one. These run on your machine. No key is needed: you are already talking to a model, so the generative steps are handed back to it and the tool does the deterministic part, which is the grading. Call skill_gate with no key and it returns the prompt per spec; you attempt them and call again with attempts. Call skill_forge with your own candidate and usage and it grades both. Set ANTHROPIC_API_KEY if you would rather a model be called for you; either way your context, your specifications and your data never reach us. Only a finished skill is published, and only when you call skill_publish.

Use them in that order, and expect skill_gate to say no. Of thirteen contexts we measured, nine were refused because the cheap model already did the job. A refusal costs cents and saves you from adding something nobody needs.

skill_check_checker is not a formality. A checker that accepts a broken file reports a perfect score for a model that did nothing; a checker that demands what the task never asked for invents failures and makes any skill look like a triumph. We wrote three versions of one checker in an hour and two of them did exactly that. Neither was visible by reading it.

The older names rung_search, rung_get, rung_install, rung_verify and rung_bank still work and do the same things.

Using it

Ask your agent for something it usually fumbles. A harness with this server installed will find the skill first:

skill_search(q="fillable pdf form")
skill_install(id="rec_0b825e...", dir="~/.claude/skills")

skill_install writes SKILL.md and the code into a folder, and your harness loads it like any other skill. Nothing runs on our servers: the skill executes wherever your model does.

Or without this package at all, since the API is public:

curl "https://api.neruva.io/v1/commons/skills?q=combo%20chart"
curl "https://api.neruva.io/v1/commons"          # every endpoint, described

Building and publishing one

Publishing is free and always will be. What you get back is the part you cannot give yourself: your code put through an exact check on cases it has never seen, signed if it passes, hosted at a name you own, and reachable by any agent.

The whole loop, from a job your model keeps getting wrong to a published skill:

skill_check_checker(checker_code=..., specs=[...])   # can it be broken? if not, fix it
skill_gate(context="...", checker_code=..., specs=[...])   # is a skill even warranted?
skill_forge(context="...", checker_code=..., specs=[...])  # build it

# or, once you have code and a guide, everything above in one call:
skill_certify(context="...", checker_code=..., specs=[...], code=..., doc=...,
              usage=[{"spec_id": "s3", "code": "..."}, ...])   # your own scripts using it

skill_publish(
  name="invoice_total_check",
  skill="pdf", task="pdf-generation", language="python",
  purpose="Make a purchase order whose line items add up to the total shown.",
  description="A purchase order PDF whose line items total correctly.",
  code=..., doc=..., entry_points=["build_purchase_order"],
  onecall={"wrapper": ..., "schema": {...}, "example": {...}},
  evidence=["16 held-out orders, exact checker: 0.31 alone, 1.00 with this"],
)

The certificate skill_certify returns has two lists, checked and not_checked. Read the second before quoting the first. A checker gated on spoilers alone is known to reject broken output and not known to accept correct output built another way; three held-out specs is a smoke test, not a rate. The certificate says so instead of leaving you to find out.

Two fields decide whether anyone ever uses what you publish. purpose is one plain sentence saying what it is FOR, in the words someone types when they need it: without it your skill is indexed on its own specification and nobody searching in plain language will find it. onecall is a single build() entry point plus a JSON schema, so an executor fills in arguments instead of recovering your function signatures from prose.

Also in this package

An account on Neruva carries a private memory and replay surface: records, recall, snapshots and audit. Those tools ship here too, but they are off by default. What you get on a plain install is the six skill tools and nothing else, because a long tool list makes a model choose worse and those tools do nothing without an account.

Turn them on if you have one:

{
  "mcpServers": {
    "neruva": {
      "command": "neruva-mcp",
      "env": { "NERUVA_API_KEY": "nv_...", "NERUVA_MEMORY": "1" }
    }
  }
}

Licence

Apache 2.0. Built by Clouthier Simulation Labs, Ontario, Canada.