runx

runx turns agent work into something you can run, govern, and audit. A "skill" is the human-facing unit of delegated work; underneath, skills compose into execution graphs. runx owns the machinery around them: authority propagation, scope attenuation, policy enforcement, approval routing, and linked receipts across the whole run.

The core is a trusted Rust local runtime that works without Node, Python, or any language toolchain installed. You author a skill as a portable SKILL.md file with a colocated X.yaml execution profile, then invoke it from the CLI, an MCP tool surface, an IDE plugin, or a source-event trigger. Every path normalizes into the same execution envelope, so behavior, policy, and receipts stay consistent no matter how a run is kicked off.

Skills can call out to any model provider (OpenAI, Anthropic, and others) for managed agent work, while deterministic tools, human approvals, and required inputs keep their local, reproducible behavior. Local skill execution can be sandboxed by declared policy (cwd scope, environment allowlists, network intent, writable paths), and every run emits an append-only, signed JSON receipt. runx history verifies those receipts and reports each as verified, unverified, or invalid, giving you a tamper-evident record of what an agent actually did.

runx is the generic engine, not a single product workflow. Teams ship their own capability packs (skills, runners, and tools) inside their repos and run them through normal skill invocation. A public registry distributes skills across three trust tiers (first-party, verified, community), and npm create @runxhq/skill scaffolds standalone community packages.

Good fit for: governed AI automation, agent orchestration with audit trails, reproducible CI/agent pipelines, MCP server hosting, and teams who want delegated AI work that is policy-bound and verifiable rather than a black box.