Mesh Decisions Playground

A small local web UI for text → question → decision using Laya through Mesh LLM. Ask a Yes/No question or choose between 2–16 labels, then inspect probability bars and the original API response.

Screenshot uses deterministic test data, not a benchmark.

Run

Requires Node.js 22+ and mesh-llm v0.77.0+ installed separately. Start Laya in one terminal:

mesh-llm serve --model meshllm/laya-multilingual-F16-GGUF

Mesh downloads the compatible GGUF and reuses its cache. Laya defaults to CPU. Wait until the model is ready. In another terminal:

git clone https://github.com/Mesh-LLM/mesh-decisions-playground.git
cd mesh-decisions-playground
npm start

Keep npm start running, then open http://127.0.0.1:8787 in a browser on the same computer. This is a local web app, not a hosted website. With Node.js installed, no npm install or build step is needed to run the playground.

To use another Mesh endpoint or playground port:

MESH_URL=http://127.0.0.1:9447 PORT=8788 npm start

MESH_URL is an origin, without /v1, credentials, or query parameters. The app only proxies GET /v1/models and POST /systemone. It never starts, stops, downloads models for, or changes settings on your Mesh node.

You can co-host a chat model using Mesh's repeated --model option:

mesh-llm serve \
  --model meshllm/laya-multilingual-F16-GGUF \
  --model Qwen/Qwen2.5-0.5B-Instruct-GGUF:Q8_0

Model discovery and limitations

Only explicit IDs advertising system_one appear in the picker. auto and mesh are excluded; architecture/name alone never establishes support.

Start by connecting directly to the node serving Laya. Mesh v0.77.0 has been observed listing remote Laya without its capability, and returning 404 for a remote decision. This app does not repair mesh routing or promise remote support. If no models appear, check the chosen node's /v1/models, model readiness, and MESH_URL, then click Refresh. On a 404, connect directly to the serving node. Management-console /systemone passthrough is not assumed.

Yes/No maps to type: "noul"; choices map to type: "choice" with each label also used as its description. Requests share text as state and use the named question decision. This is not an OpenAI chat interface. See the System One contract.

Probabilities are model estimates, not calibrated guarantees. There is no automatic tool execution, chat gating, or safety-screening layer here.

Privacy and boundaries

• Server binds 127.0.0.1 only. Host and Origin checks reject foreign sites.
• Upstream is configured at startup, never supplied by a browser request.
• No analytics, external browser assets, prompt logging, or persistent browser storage.
• Text is sent to the configured Mesh endpoint. That endpoint's own routing, retention and logging policies still apply. Do not send sensitive text to an untrusted endpoint.
• JSON request limit: 128 KiB; upstream response limit: 1 MiB; timeout: 30 seconds.
• No upstream authentication support. Do not expose unauthenticated Mesh publicly to make this demo reachable. This is a local demo, not a public multi-user service.

Develop and verify

npm ci
npm test
npx playwright install chromium
npm run test:browser

Server/contract tests use a real loopback mock upstream. Browser tests cover Yes/No, choices, loading, invalid choices, errors, empty discovery, refresh, HTML-safe results and a 390px mobile layout. They use fixture responses and refresh the README screenshot. This app has not yet been browser-verified against live inference; the underlying published-model setup was previously verified with released Mesh v0.77.0. Tests start and stop only their own web servers, never Mesh.