Docs

Quick start

Three steps: install the MCP tools, search the shelf, install a skill as a skill. Reading never needs a key.

1. Install the MCP tools

pip install neruva-mcp

Add the server to your MCP client. For Claude Code, the entry looks like this (a key is only needed to bank skills):

{
  "mcpServers": {
    "neruva": { "command": "neruva-mcp", "env": { "NERUVA_API_KEY": "optional, for banking" } }
  }
}

2. Find and install a skill

# in Claude Code (MCP tools from neruva-mcp)
rung_search(q="pdf_fillable_form_rungs")
rung_install(id="rec_0b825e5402a34e4da7f7c044b7d1e56d", dir="~/.claude/skills")
# the skill folder is now loaded like any other skill

Installing writes a folder with SKILL.md, the code and a small JSON with the certificate and evidence. Your harness loads it like any other skill.

3. Let your model call it

The usage doc inside the skill tells the model which entry point to call and with what. Small models do best when the skill exposes a single call with a JSON schema; most skills on the shelf come with one.

MCP tools

tool	does
skill_search	search the shelf by meaning, optional task and skill filters
skill_get	fetch one skill in full: code, guide, evidence, certificate
skill_install	write it as an Agent Skills folder into a directory
skill_tasks	the controlled list of kinds of job, for filtering a search
skill_verify	recompute the hash and check the signature server-side
skill_publish	publish your own (needs a key)

HTTP API

Base URL https://api.neruva.io

method	path	notes
GET	/v1/commons/rungs?q=&skill=&limit=	Search by meaning; filter by skill. Public.
GET	/v1/commons/rungs/{id}	Full rung: code, usage doc, evidence, entry points, certificate. Public.
GET	/v1/commons/rungs/{id}/code	The code as text. Public.
GET	/v1/commons/rungs/{id}/skill.md	The skill rendered as a SKILL.md file. Public.
POST	/v1/commons/verify	{id} or {code, cert}: recompute the hash, check the ed25519 signature. Public.
GET	/v1/commons/stats	Counts for the shelf. Public.
POST	/v1/commons/rungs	Bank a skill (needs an API key). Re-checked before listing.

FAQ

What is a skill?

A small library of helper functions for one kind of task, written by a frontier model, checked by a machine on cases it never saw, and signed. Your model calls it instead of writing the code.

Which models can use a skill?

Any model that can call a function or fill in a JSON schema. We measure with DeepSeek-chat and with local Qwen3 models down to 0.6B parameters; see the evidence page.

Does my data leave my machine?

Reading the shelf sends nothing. Running a skill is local to wherever your model runs. Forging on our side uses the task description and examples you send us and nothing else.

How do I know a skill is not tampered with?

Each skill carries a sha256 of its code and an ed25519 signature over its certificate. The verify endpoint recomputes both server-side; the MCP tool does the same.

What if the skill I need does not exist?

Forge one from the pricing page: $20 for a single skill on your task, or a plan with a monthly allowance. If it cannot be built in five rounds you get a refund and the failure report.

Can I contribute?

Yes, per skill, opt-in. We take the helper code and its usage doc, re-run the gates, and list it signed if it passes.

Support

Email kyle@simgen.dev. Bugs and requests also go to the GitHub issues.