The graph engine compiled to WebAssembly. This directory is the phase-1 spike — proof that the engine layer (compile, run, executor, scheduler, coordinator, state, channels, NodeCache) builds and executes unmodified under Emscripten.

For comparison: native NG totals 5.5 MB; the LangGraph stack (langgraph + langchain + openai + httpx + pydantic + langsmith) is 31 MB of pure-Python that does not even attempt to ship to a browser. NG fits inside L3 cache twice over and is small enough that a typical SaaS landing page already loads more JS than this engine takes.

• GraphEngine::compile(json) — JSON definition → executable engine.
• engine->run(cfg) — synchronous run with InMemoryCheckpointStore.
• Custom nodes registered via NodeFactory::register_type — leaf semantics carry over from the C++ / Python paths.
• All the v0.1.6 features compile clean: set_worker_count, set_node_cache_enabled, channels with reducers, conditional edges, Send fan-out, Command routing, interrupts.
• C++20 coroutines (asio's header-only awaitable pieces) work under Emscripten 5.0.

source /opt/emsdk/emsdk_env.sh

em++ -std=c++20 -O3 -flto -fexceptions -pthread \
  -sALLOW_MEMORY_GROWTH=1 -sPTHREAD_POOL_SIZE=4 \
  -DASIO_STANDALONE -DASIO_NO_DEPRECATED \
  -I include -I deps/asio/include -I deps/yyjson \
  wasm/smoke.cpp \
  src/core/json.cpp deps/yyjson/yyjson.c \
  src/core/graph_engine.cpp src/core/graph_compiler.cpp \
  src/core/graph_coordinator.cpp src/core/graph_executor.cpp \
  src/core/scheduler.cpp src/core/graph_state.cpp \
  src/core/graph_node.cpp src/core/graph_loader.cpp \
  src/core/graph_checkpoint.cpp src/core/store.cpp \
  src/core/provider.cpp src/core/tool.cpp \
  src/core/react_graph.cpp src/core/plan_execute_graph.cpp \
  src/core/deep_research_graph.cpp src/core/node_cache.cpp \
  -o wasm/smoke.js

Run with node wasm/smoke.js. No browser flags needed.

-pthread is required because compile() provisions a default thread_pool sized to hardware_concurrency(). Single-threaded WASM is also possible — pass -sPTHREAD_POOL_SIZE=0 and call engine->set_worker_count(1) before run().

1. 2-A — Embind JS bindings. Expose GraphEngine, RunConfig, ChannelWrite, Send, Command to JS. A JS function can register itself as a node implementation; the engine calls back into JS for each node execution. Estimated 1-2 days.

2. 2-B — fetch-based HTTP transport. Provide a transport interface that SchemaProvider consumes; the WASM build wires it to fetch(). Same provider code targets either backend. Estimated 3-5 days.

3. 2-C — npm package. Publish as @neograph/wasm so apps can npm install the engine + JS bindings without their own build. Estimated 1-2 days.

After Phase 2 the engine can run Originator-issued graphs entirely in a browser tab — leaves call BYOK Anthropic / OpenAI / Bedrock keys via fetch(), transformers.js / built-in AI for local inference, and results flow back through channels into a Result Envelope. That's the runtime side of the NeoProtocol Executor role.