Scott built Skillscript to solve a recurring problem with his NanoClaw agent: every morning brief required the model to re-derive the same fixed procedure, burning frontier tokens and drifting slightly each run. The language lets a frontier model (or human) author a specific, auditable procedure once, then hands execution to a cheap local model — currently Ollama — for repeated autonomous runs. Version 0.30 is MCP-native and self-hosted via `npm install -g skillscript-runtime`.

Skillscript deliberately constrains expressiveness: no `eval`, no arbitrary imports, no subprocess, no unbounded loops. Skills are declarative DAGs of typed dispatches — calls to tools, models, and data stores through swappable connector contracts. Computation lives in connectors; coordination lives in skills. Safety is enforced at the language level: sandboxed grammar, connector-mediated capability, static validation before admission, and Ed25519 signed approval for effectful operations. The runtime never holds the signing key.

A skillscript skill compiles into the same artifact shape as an Anthropic/OpenAI Skill — a `# Skill: <name>` header with instructional markdown body — so it loads into existing agent contexts. Mature deployments use both: Skills as agent-facing capability descriptions, skillscript as the higher-leverage authoring layer underneath. The grammar disciplines both human and agent authors into structure, not prose.

The project is pre-1.0 with acknowledged rough edges: first-run setup friction, grammar still moving, Ollama assumption for local models. But the architecture makes a clear bet: routine orchestration should be cheap, inspectable, and crystallized, not re-reasoned at inference time. If the connector ecosystem matures, Skillscript could become the standard substrate for persistent agent capability.