Execution Surfaces

Execution Surfaces

MLPL is one language exposed through several clients and deployment shapes. A surface is how a user interacts; a backend is where computation happens.

flowchart TD
    User[User] --> WebLocal[Browser playground]
    User --> CLI[Native REPL or script]
    User --> WebRemote[Connected browser]
    User --> Rust[Embedded or compiled Rust]
    WebLocal --> WASM[In-browser WASM evaluator]
    CLI --> Local[Local evaluator]
    WebRemote --> Serve[mlpl-serve session]
    Rust --> Native[Native runtime subset]
    Local --> Devices[CPU, MLX, or CUDA]
    Serve --> Devices

Loading

Surface	State lives in	Main strengths	Main limits
Browser-only playground	Browser memory	No install, tutorials, inline UI and visualization	No arbitrary local files, host processes, or direct native GPU
Native REPL	Local process	Full REPL, files, traces, scripts, optional native GPU	Single local client and terminal-oriented output
Script execution	Local process	Repeatable .mlpl programs and demos	Non-interactive unless instrumented
Connected terminal	Server session	Remote compute and persistent shared state	Depends on server endpoints and network
Connected browser	Server session	Browser UI with native CPU/GPU compute	Must configure connection, auth, CORS, and device availability
mlpl! embedding	Rust host process	MLPL expressions inside Rust	Supported compilation subset
mlpl build	Native binary	No parser/interpreter in result	Supported lowering subset, fewer dynamic features

Browser-only versus connected browser

These look similar but are operationally different. Browser-only evaluation happens inside WASM on the user's CPU. Connected evaluation sends programs to mlpl-serve; that server may have filesystem access, persistent state, MLX, or CUDA. The UI should probe /v1/devices and only enable demos supported by the connected build.

Local versus service-backed accelerators

MLX can be compiled in-process and also has a peer-service design. CUDA currently has an in-process connected-server vertical slice: the CUDA machine runs mlpl-serve --features cuda, and clients send work to it. Whole device-scoped programs reduce network chatter; device tensors should remain remote until explicit materialization where the peer-handle path is used.

Continue with Capability Matrix to compare features, or select Browser and WASM, Native CPU, Apple Silicon and MLX, or NVIDIA and CUDA.