LM Studio Research UI

A network-accessible research interface for LM Studio. Run it once on the machine hosting LM Studio; access it from any browser on your network. Designed for maximum observability — every token, timing metric, and raw API payload is visible in real time.

---

Prerequisites

Requirement	Version
Node.js	≥ 18.0.0 (uses native fetch and ReadableStream)
LM Studio	Any version with the local server enabled
Browser	Any modern browser on the same network

---

Installation

git clone <this-repo>
cd harness-this
npm install

---

Running

Basic (LM Studio on the same machine, default port)

npm start

The server starts on http://0.0.0.0:3000 and proxies to http://localhost:1234.

Development (auto-restart on file changes)

npm run dev

With environment variables

All configuration is done via environment variables — no config files to edit.

Variable	Default	Description
PORT	3000	Port the UI server listens on
LM_HOST	localhost	Hostname or IP of the machine running LM Studio
LM_PORT	1234	Port LM Studio's local server is bound to

Examples:

# LM Studio running on a different machine
LM_HOST=192.168.1.42 npm start

# LM Studio on a non-default port
LM_PORT=8080 npm start

# Everything custom
PORT=8000 LM_HOST=192.168.1.42 LM_PORT=1234 npm start

Accessing from another machine

Find this machine's local IP:

ipconfig getifaddr en0       # macOS (Wi-Fi)
ipconfig getifaddr en1       # macOS (Ethernet)
ip route get 1 | awk '{print $7}' # Linux

Then open http://<that-ip>:3000 in any browser on the network.

The server binds to 0.0.0.0 (all interfaces) by default, so no extra firewall configuration is required on most setups.

---

Making sure LM Studio is ready

1. Open LM Studio
2. Go to the Local Server tab (the <-> icon in the left sidebar)
3. Click Start Server — the default port is 1234
4. Load at least one model

The UI will show a green status dot and populate the model dropdown when it successfully connects.

---

Using the Interface

Layout

┌──────────────────────────────────────────────────────────────────┐
│  ⊞  ◉ LM Research  ●  [Model dropdown]    ⊕ New  ↓ Export  ⊡   │
├─────────────┬──────────────────────────────┬─────────────────────┤
│  PARAMETERS │                              │  Console │Stats│Raw │
│             │   Chat messages              │                     │
│  Preset     │   stream here                │  Live server events │
│  Sliders    │                              │  broadcast via SSE  │
│  System     │                              │                     │
│  Prompt     │──────────────────────────────│                     │
│             │  [token count]  [Markdown ✓] │                     │
│             │  ┌──────────────────────┐    │                     │
│             │  │  Type message here…  │[→] │                     │
└─────────────┴──────────────────────────────┴─────────────────────┘

Left panel — Parameters

Control	Range	Description
System Prompt	text	Injected as the system role message on every request
Preset	Default / Creative / Precise / Code	One-click parameter bundles (see table below)
Temperature	0 – 2	Randomness of token sampling
Top P	0 – 1	Nucleus sampling cutoff
Top K	0 – 200	Limits vocabulary to top-K tokens at each step
Repeat Penalty	0.5 – 2	Penalises recently used tokens
Max Tokens	-1 – 32768	Maximum completion length (-1 = model default)
Seed	-1 – 2147483647	Fixed seed for reproducible outputs (-1 = random)
Stop Sequences	one per line	Generation stops when any sequence is matched

Presets:

Preset	Temp	Top P	Top K	Rep. Penalty	Best for
Default	0.7	0.95	40	1.1	General use
Creative	1.2	0.98	80	1.05	Brainstorming, fiction
Precise	0.1	0.90	20	1.15	Factual Q&A, structured output
Code	0.15	0.95	20	1.2	Code generation, debugging

Center panel — Chat

• Type in the input box and press Ctrl+Enter (or ⌘+Enter on Mac) to send, or click Send.
• Responses stream token-by-token with a blinking cursor.
• Each assistant message shows a stats footer: token count, generation speed, time to first token, total time, and finish reason.
• Copy button on each message copies the raw text to clipboard.
• Copy code button appears on every code block.
• ↺ Retry button re-runs the last assistant turn with current parameters (useful for seed/temperature experiments).
• Toggle Markdown in the toolbar to switch between rendered and raw text.
• Esc or the ■ Stop button aborts generation mid-stream.

Right panel — Debug

Console tab

Real-time feed of every server-side event, broadcast via Server-Sent Events:

Badge	Color	Meaning
REQ	Blue	Outgoing request to LM Studio
RES	Green	Response received from LM Studio
ERR	Red	Any error (network, parse, upstream)
SYS	Purple	Server lifecycle events (client connect/disconnect)
STAT	Yellow	Per-generation summary line after each completion

Click ✕ to clear the console log.

Stats tab

Detailed metrics for the last generation and session totals:

Metric	Source	Notes
Prompt tokens	LM Studio usage field	Available when include_usage is supported
Completion tokens	LM Studio usage field	Falls back to stream chunk count
Tokens / sec	Calculated	completion_tokens ÷ (elapsed − TTFT)
Time to first token (TTFT)	Measured by proxy	Latency from request sent to first content chunk
Total time	Measured by proxy	Wall-clock time for the full generation
Finish reason	LM Studio	stop (clean end), length (hit max_tokens), other

Raw tab

Full JSON of the last Request (sent to LM Studio) and Response (all streamed chunks) with syntax highlighting. Use this to inspect exactly what parameters were sent, verify stop sequences, or copy payloads for use in other tools.

---

Keyboard shortcuts

Shortcut	Action
Ctrl/⌘ + Enter	Send message
Esc	Abort current generation
Ctrl/⌘ + L	Focus the input box

---

Exporting conversations

Click ↓ Export in the header to download a JSON file containing:

• Model used
• System prompt
• All parameter values at time of export
• Full message history with per-message stats
• Session totals

File is named lm-research-<timestamp>.json.

---

Troubleshooting

Red status dot / "Cannot reach LM Studio"

• Confirm LM Studio's local server is started (the <-> tab, click Start Server)
• Check LM_HOST and LM_PORT match the LM Studio server address
• Make sure no firewall is blocking port 1234 between the two machines

Model dropdown shows "No models loaded"

• LM Studio is reachable but no model is loaded — load a model in LM Studio first

Responses appear cut off

• Increase Max Tokens in the parameters panel, or click ↺ Retry after raising the limit

Generation is slow

• Watch the Tokens / sec value in the Stats tab — this reflects the model's throughput on your hardware, not the UI
• Lower Top K (e.g., 20) can slightly improve speed on some backends