Generative UI Suite

Agents publish interactive HTML surfaces — real Web Components, not screenshots — by writing files into a directory. A Rust server discovers them (no submission API), serves the live version of each surface, and writes user interactions (backflow) back into the same directory, where the producing agent reads them. The filesystem is the bus; discovery and backflow are symmetric.

It re-works the core idea of ui.plan.ai — make agent-generated UI work inspectable — but swaps pictures+click-zones for real HTML, the submission API for a directory schema, and one-way events for a directory-based backflow loop.

Concepts

• agent → deck → surface → version. A deck is a navigable set of interconnected surfaces (main by default). A surface is a regenerable "dashboard" whose newest version is live; history is retained.
• backflow: every interaction is written to the surface's events/ directory.

artifacts/{agent}/{deck}/{surface}/v/{version}/{frame.json,index.html}
artifacts/{agent}/{deck}/{surface}/events/{unix-ms}-{action}.json
artifacts/{agent}/{deck}/deck.json     # optional: entry surface + nav order

Quick start

Everything runs through the Nix flake (pinned Rust + node, cross-platform):

nix develop                                   # dev shell
cargo run -p gui-server -- --artifacts ./artifacts
# open http://localhost:8080/ for the inspector, or a surface directly:
# http://localhost:8080/demo/sales/
# http://localhost:8080/demo/handlers/   sandboxed JS/Lua/WASM handlers (click Bump → counts)

Surfaces can ship sandboxed handlers (JS/Lua/WASM) that run server-side on a backflow action — they fold a per-surface KV view (bound in the UI via data-kv) and define the response. Handlers run by default (default-deny capabilities + resource limits); pass --disable-handlers to turn execution off.

Build a single binary without the shell:

nix build .#gui-server && ./result/bin/gui-server --artifacts ./artifacts

Verify the loop end to end

nix develop                                   # pinned Rust + node dev shell
cargo test                                    # all crates
node --test suite/*.test.js                   # JS runtime helpers

# Serve the committed demo deck and open it:
cargo run -p gui-server -- --artifacts ./artifacts
#   http://localhost:8080/            inspector (live)
#   http://localhost:8080/demo/sales/ real interactive HTML; nav to region-detail/settings

# Drive the loop from an agent (in another shell, same artifacts dir):
cargo run -p example-agent -- --artifacts ./artifacts wait sales   # blocks
# …click "Refresh" in the browser → an event file lands in
#   artifacts/demo/main/sales/events/ and the agent prints it.

# Or without a browser, via the CLI:
cargo run -p gui-cli -- --artifacts ./artifacts backflow emit demo sales --action refresh
cargo run -p gui-cli -- --artifacts ./artifacts index

The MCP server (cargo run -p gui-mcp -- --artifacts ./artifacts) exposes the same authoring/awaiting flow to any MCP-speaking agent (publish_surface, define_deck, wait_for_backflow, …).

Direct access — an agent that authors the whole app via the MCP

The direct-access example runs the HTTP server, the gui-mcp MCP server, and a swiftide LLM agent in one process (separate libraries, wired in-process over an in-memory pipe — no subprocesses). The agent connects to the MCP server, loads the authoring-ui skill as its system prompt, and builds the app itself by calling the MCP tools (publish_surface, define_deck, …).

cargo run -p direct-access -- --artifacts ./run --mock        # deterministic, offline
# open http://127.0.0.1:8080/  — the surfaces the agent publishes appear live

# Real / OpenAI-compatible backend (opt-in; custom base URL supported):
OPENAI_API_KEY=… cargo run -p direct-access --features openai -- \
    --base-url https://your-endpoint/v1 --model gpt-5.4-mini \
    --goal "Build a signup screen, then a thank-you screen"

Notes:

• The default build uses a built-in mock LLM (no network) that drives one publish_surface call — what the test exercises. The real backend is gated behind --features openai (uses swiftide from git + async-openai 0.33; needs a recent cargo for git swiftide's 2024 edition).
• Pass a --model your endpoint hosts (query GET {base-url}/models). The agent is generic over any swiftide ChatCompletion, so the backend is pluggable.
• Verbose logging is on by default (stderr); tune with RUST_LOG.
• Verified live against a custom base URL: the agent published a real <ui-*> component tree through the MCP.

Crates

Crate	Role
gui-schema	The on-disk contract: manifests, decks, events, handlers, component tree → <ui-*> HTML. Generates schemas/*.json.
gui-agent	Agent API: publish surfaces (+ handlers), define decks, read/wait on backflow.
gui-server	Discover + serve surfaces, route backflow, run handlers, SSE, inspector.
gui-runtime	Sandboxed handler engines (JS/Lua/WASM) behind one capability-limited host API.
gui-mcp	rmcp stdio MCP server — publish_surface, define_deck, wait_for_backflow, …
gui-cli	gui debug CLI — index, validate, deck, surface, publish, backflow, watch.

See AGENTS.md for the agent-facing reference and hard rules.

License

MIT.