Skillscript Review 2026

A declarative, sandboxed, non-Turing-complete language and runtime that lets AI agents crystallize learned procedures into auditable, reusable skills.

Skillscript is an open-source language and runtime built on a simple observation: AI agents re-derive routine tasks from scratch every time, burning expensive frontier-model inference on procedures with a known shape. Skillscript gives agents a substrate to write themselves down in — a skill is a small declarative program, a dependency DAG of typed operations, that an agent authors once and the runtime fires many times, on cron, on events, or on demand. The language is deliberately not Turing complete: it cannot eval, cannot spawn subprocesses, cannot import arbitrary packages. Skills can only do what configured connectors permit, which is the safety story for code written by agents rather than humans. Skills are statically linted before admission, and in secured mode only skills carrying a valid Ed25519 operator signature can perform effectful operations — an unapproved or tampered skill is inert. MCP support is central and two-way: the runtime exposes its full surface as an MCP server (skill_list, skill_write, lint_skill, compile_skill, execute_skill, triggers, health metrics, and more), so any MCP-speaking agent can discover, author, validate, and run skills entirely over the wire; and skills themselves dispatch work to LLMs and tools through MCP-backed connectors, letting a frontier model delegate bounded sub-tasks like classification and extraction to cheap local models. Skills come in three shapes: headless (cron monitors and batch jobs), augmenting (context injected when an agent wakes), and template (recipes an agent executes). Install with npm install -g skillscript-runtime. MIT licensed, pre-1.0.