Rotifer MCP Server

What rotifer-protocol/rotifer-mcp-server does

Rotifer Genes are reusable capabilities that can be discovered, evaluated, installed, composed, compiled, tested, run, and published through MCP. The server covers both Rotifer Cloud data and the current project workspace, allowing an AI coding client to move from finding a Gene to creating an Agent pipeline around it.

Cloud-facing operations search the Gene registry, retrieve phenotype and metadata, inspect download and reputation data, compare fitness metrics, view Arena rankings, and examine creator profiles or version histories. Local operations inspect genes/ and .rotifer/agents/, install registry artifacts, create new Gene projects, wrap existing functions or SKILL.md files, and manage compiled or executable Genes.

How it works

The server uses MCP stdio transport and can be launched with npx @rotifer/mcp-server. MCP clients invoke tools for discovery, lifecycle management, Agent composition, authentication, and analytics. The README also defines MCP resources for Gene details, statistics, creators, leaderboards, and local inventories, along with prompts for guided creation, architecture, protocol orientation, and Arena evaluation.

Local Gene commands delegate to the Rotifer CLI. Compilation uses the TypeScript-to-WASM toolchain, while testing and execution can use the sandbox. doctor checks the esbuild and javy setup without changing the environment. Arena submissions calculate fitness by running the local Gene rather than accepting caller-supplied scores.

Arena results distinguish verified measurements from results still under evaluation or not eligible for ranking. Rankings include fitness and related measurements such as safety, success rate, latency, resource efficiency, sample size, and caller count.

Setup and configuration

Add the following MCP server entry to an MCP client configuration:

{
  "mcpServers": {
    "rotifer": {
      "command": "npx",
      "args": ["@rotifer/mcp-server"]
    }
  }
}

Node.js 20 or newer is required. Cursor and Claude Desktop use the same stdio configuration, and the command also applies to other MCP clients that support stdio transport. Log in through the login tool before using arena_submit or publish_gene; OAuth authorization is handled through GitHub or GitLab and credentials are stored locally.

Tools and capabilities

• Discover Genes by query, domain, fidelity, sort order, and page; inspect details, versions, statistics, reputation, and Arena rankings.
• Compare two or more Genes using side-by-side fitness breakdowns.
• List local Genes and Agents, including metadata, compile state, genome composition, strategy, and reputation.
• Initialize, scan, wrap, compile, test, run, install, publish, and roll back Genes.
• Create and execute local Agents composed from one or more Genes.
• Check the local compilation toolchain with doctor.
• Authenticate with Rotifer Cloud and view personal or server call analytics.

Publishing validates a Gene with a security scan, requires the required security grade, and can optionally submit it to the Arena. install_gene verifies the declared artifact size and SHA-256 when a compiled WASM artifact is available.

Limitations and notes

Genes used for Agent creation, execution, compilation, testing, and Arena submission must exist in the local workspace where the relevant operation requires them. Agents are stored under .rotifer/agents/, and local Genes are discovered from the project’s genes/ directory. Rollback restores only the most recent overwritten copy; each snapshot is consumed after one rollback.

Arena submission and publishing require Rotifer Cloud authentication. Arena ranking data should be interpreted using its tier: verified measurements are the preferred basis for choosing a Gene, while estimated, disqualified, or not-yet-corroborated results do not provide the same ranking confidence.