Blog
AI agent tools run with your permissions. They don't have to.
Every MCP server you install executes as your user, with your files and keys in reach. ACT packages tools as WebAssembly components with a declared capability ceiling, explicit grants and an audit trail. Live demo: real pandas in your browser tab, touching one file.
WASI Python in 2026: dlopen works, NumPy still doesn't
A self-hosted Python sandbox for AI agents that can't import NumPy — and the surprise is that the blocker isn't the runtime anymore. A field report on the state of WASI Python in May 2026: the runtime is ready, PEP 816 just went Active, and the wheels are still ahead of us.
ACT 0.7: sessions, three production bridges, and auth-via-args
Components can now own per-session state through `act:sessions/session-provider` — a small WIT interface for components that need to hold a database connection, a parsed OpenAPI spec, or an upstream MCP handshake. Three bridges in `actpkg` are rebuilt around it: act-http-bridge, mcp-bridge, openapi-bridge. Auth moves into open-session args, where it belongs.
A tour of actpkg — components on the protocol
A protocol is only interesting if people actually write for it. Here's what's published on actpkg today — storage, network bridges, utilities, language embedders — and what each one is for.
The capability ceiling — how ACT sandboxes third-party tools
ACT layers a declaration-as-ceiling policy model on top of WASI's capability imports and wasmtime's VM isolation. The declared ceiling, runtime policy, DNS-layer deny-CIDR, and per-hop redirect re-check combine to make a tool from an untrusted author safe to run under your agent.
MCP servers, sandboxed — introducing ACT
Today's MCP servers ship as `npx`, `uvx`, or `curl | bash` — ambient-permission native processes with your full user access. ACT packages tools as sandboxed WebAssembly components with declared capabilities and deny-by-default filesystem and network, so an agent can safely call tools written by a stranger.