Muse Glimmer — local agent console

1 · Connect your local runtime

This page talks to http://localhost. Browsers allow that from an HTTPS page because loopback is a trustworthy origin — but your runtime still has to permit the cross-origin request. Nothing is sent anywhere else; there is no server behind this site to send it to.

2 · Capability tracks

Each track is a scripted scenario over the local workspace, scored by reading the run's action ledger — which tools were called, with what arguments, and what changed — rather than by asking a model to grade the prose. You can read every prompt before running anything.

3 · Protocol inspector

No model required. This is the exact byte string this page would send to a completion endpoint for the selected track — channel tokens, tool preamble, ATEM examples and all. It is generated by the same code path the agent loop uses.

Why the wire format is the interesting part

Muse Glimmer does not emit OpenAI-style JSON tool calls. Its chat template builds a channel-token transcript and asks for calls in a dialect the template itself calls ATEM:

<|start|>assistant to=self<|message|>Search first, then escalate.<|eom|>
<|start|>assistant to=workspace.search_tickets<|message|><atem:function_calls>
<atem:invoke name="workspace.search_tickets">
<atem:parameter name="query">billing</atem:parameter>
<atem:parameter name="limit">20</atem:parameter>
</atem:invoke>
</atem:function_calls><|eot|>

Three consequences that cost real debugging time:

• Reasoning is a routing target, not a tag. Chain-of-thought goes to to=self and closes with <|eom|>; the visible answer goes to to=user and closes with <|eot|>. There is no <think> block to strip. In practice the model does not reliably emit that closing <|eom|> before switching recipients — I hit this on the real weights, and upstream llama.cpp is patching the same class of bug.
• ATEM is not XML, and the template says so. Verbatim: "The output is not expected to be valid XML and is parsed with regular expressions." Parameter bodies carry raw <, & and unbalanced quotes, so an XML parser rejects valid calls.
• Everything arrives as text. ATEM has no types — true and "true" are the same bytes. The JSON Schema you supplied is the only thing that can tell them apart on the way back, so a scaffold that ignores the schema hands your tools strings where they expect booleans and integers.

A trap worth naming. The tool preamble the template injects into the system message contains two complete, syntactically valid ATEM blocks — the $FUNCTION_NAME syntax sketch and the example_tool_name worked example. Anything that scans a whole prompt for tool calls, rather than only the completion, will report two calls the model never made. The conformance suite pins this as a test so it cannot regress into a surprise.

Conformance against the real chat template

The renderer in protocol/atem.js is a hand-written port. To keep it honest, every fixture is also rendered by the actual Jinja template shipped in meta-models/Muse-Glimmer-30B, under the same environment transformers.apply_chat_template builds, and the two outputs are compared byte for byte.

Regenerate with python3 conformance/render_reference.py && node conformance/run.mjs.

Setup

You need the weights and a server. Roughly 17–20 GB on disk, 24 GB of VRAM for the smaller quant.

Ollama

ollama pull muse-glimmer:30b-q4_
# CORS: the page's origin must be allowed
OLLAMA_ORIGINS="https://carlosrymer.github.io" ollama serve

llama.cpp

llama-server -hf meta-models/Muse-Glimmer-30B-GGUF \
  --host 127.0.0.1 --port 8080 \
  --spec-type draft-dflash --spec-draft-n-max 15   # optional: DFlash speculative decoding

LM Studio

Load Muse-Glimmer-30B-GGUF, start the local server on port 1234, and enable CORS in the server tab.

Sampling follows the model card: temperature 1.0, top_p 0.95, top_k 64. Reasoning strength is set through the system prompt, which is why this page renders the prompt itself rather than delegating to the runtime.