ACT

ACT, or Agent Component Tools, is a hardened AI toolchain for building sandboxed components for AI agents. It is aimed at environments where a simple local run is not enough, and it addresses the problem of third-party tools arriving with broad access, opaque provenance, and no declared execution bounds.

Its components are built as WebAssembly-based units and described as self-documenting. The workflow shown on the site includes writing a tool in Rust, Python, Go, C++, or any language that compiles to WebAssembly, compiling it into a single .wasm component, and then pushing it to an OCI registry. The component format can include schemas, descriptions, examples, and agent skills baked into the binary, and the published artifact carries signed provenance through a GitHub Actions attestation linking the artifact to its source commit. The system also uses deterministic builds so the same SHA256 can be produced on every machine.

ACT describes its runtime model as sandboxed by wasmtime with zero ambient access, meaning no filesystem access, no outbound network, and no syscall surface unless the operator explicitly grants it. It also uses a capability ceiling: components declare what they need at build time, operators grant what they are willing to allow, and the host enforces the intersection. The site says components can run over MCP, CLI, HTTP API, or in the browser, and the example command line shows a pull-and-run flow from an OCI registry. The toolchain is presented as suitable for AI agents and for humans reading the embedded documentation without running the component first.

The page includes npm installation commands for @actcore/act and @actcore/act-build.