Runx

What runx does

The runx MCP server provides an agent-facing path into Runx, a governed runtime for portable skills described by SKILL.md files. A skill contains the method, evidence requirements, decisions, and tool needs for a piece of work. Skills can pass outputs to other skills, allowing an agent or project-owned graph to compose research, drafting, release, analytics, or communication workflows.

Runx places an execution boundary around those workflows. It admits calls under explicit authority, delivers credentials at the tool boundary, supervises execution, and records a receipt for each run. The runtime, execution state, and receipt signing can run locally. A hosted registry, credential custody, and notary services are optional.

How it works

An agent first discovers a suitable skill, then examines its inputs and requirements before choosing which tools and accounts to bind. The run proceeds through open turns, so missing credentials, evidence, or decisions remain explicit requests rather than being silently assumed. Results can become inputs to later skills.

Skills may be resolved from a local directory, a URL, or the registry. Local packages take precedence when resolving a bare name. For reproducible execution, use an exact version and digest. A completed run can return a sealed status, a result, a closure disposition, and a receipt identifier. The CLI can read the run history and verify the returned receipt.

The runx MCP server is one of the interfaces available alongside the CLI and language clients.

Setup and configuration

The native CLI can be installed from npm with npm install --global @runxhq/cli, or through the standalone installer on macOS and Linux:

curl -fsSL https://runx.ai/install | sh

The smallest repository example requires Git and Node.js 20 or newer, but does not require a Runx account, model, or provider key:

git clone --depth 1 https://github.com/runxhq/runx.git
cd runx
runx skill ./examples/hello-world --run-id first-run -i message="Hello, Runx." --json

Agents can also be given the Runx operating manual at https://runx.ai/SKILL.md. Native skill exports are available for Claude Code and Codex.

Tools and capabilities

The runx MCP server and the related runtime support these documented capabilities:

• Discover skills from local paths, URLs, or the registry.
• Inspect skill inputs and requirements before execution.
• Bind selected tools and credentials at the tool boundary.
• Execute skills individually or as connected workflows.
• Pass evidence, drafts, decisions, and prior results between skills.
• Read run records with runx history and verify receipts with runx verify.
• Use exact versions and digests for repeatable skill resolution.

The repository examples include skills for operations review, lead enrichment, release preparation, content packaging, message planning, and product analytics. These are examples of compositions rather than a complete list of runtime tools.

Limitations and notes

Runx governs execution, but the agent or orchestration layer still supplies judgment. A run can stop when required accounts, evidence, approvals, or decisions are missing. The README also demonstrates stopping before publishing or sending, so externally consequential actions should remain subject to the workflow's approval requirements.

It explicitly mentions native exports for Claude Code and Codex, but client-specific MCP setup instructions are not provided. Hosted registry, credential custody, and notary components are optional; local execution does not imply that external provider accounts are configured.