LM Manager Pro Web

Display

ThemeDark mode is the default. Light mode works but some colors are more saturated than intended — designed primarily for dark.

Inference

Max tokensMaximum tokens per agent response. Higher = longer answers but slower. 8192 is a safe default for most models.

Request timeoutSeconds to wait before marking a request as failed. Increase for large models that are slow to load.

Default temperatureCreativity/randomness fallback for models with no per-model temp set. 0 = deterministic, 1 = chaotic. Most tasks work best between 0.3–0.7.

Enable reasoningSends reasoning: "on" to the LM Studio v1 API. Models that support extended reasoning (DeepSeek R1, Qwen3, QwQ) will activate it. LM Studio silently ignores this for models that don't support it. Increases quality on hard tasks but costs more time and tokens.

Auto-check machines on launchWhen enabled, the app runs Check All every time the page loads. Leave this off if you want faster startup or often open the app away from your LM Studio machines.

Director

App nameThe name shown in the topbar and browser tab. Cosmetic only.

Director nameReplaces the default director label in every auto-generated persona prompt. Set this to your own name so models address you correctly during interviews and group chat.

Global system suffixA sentence appended to every auto-generated prompt. Use this for project-wide context every model should know — e.g. the tech stack, the app name, your coding style. Only applied to non-overridden prompts.

LM Link · Device Map

Paste the output of lms ls --json from any linked machine. The app will detect each device hash and let you assign it to a machine.

Run lms ls --json onWhich machine did you run this command on? Models with no device hash belong to it.

Check All machines then open the new models modal

Roster

Default machinePre-selected in the Add Engineer modal. Set this to whichever machine you add models to most often.

Auto-generate namesParses model IDs to fill First (family name) and Middle (param count) for everyone. Won't overwrite fields you've already set. Use "Force all" in the confirm dialog to overwrite everything.

Sync model data from LM StudioUpdates file size and param count for every roster entry using live metadata from the last Check All. Run "Check All" first. Names and roles are untouched.

Roles

Roles are built from three layers. Status sets color saturation (Unhired = neutral gray, Flagship = full color). Primary specialization sets the color family and main prompt line. Secondary specializations add focused function lines on top.

Status × Primary

Secondary specializations · stackable · each adds a prompt line

Custom roles

Interview

Auto-clear history on model swapWhen switching between models in the interview tab, wipes their conversation history so each starts fresh. Useful when running the same prompt across the whole roster.

1 · Opening · auto from Director settings above

2 · Product · what you're building

3 · Stack · tech environment

4 · Team · auto-generated from roster

5 · Closing ask · what you want to learn

Preview

Group Chat

Max roundsHow many times the full team speaks before the run ends. 1 = everyone speaks once. Increase for debate or iterative refinement tasks.

History depthHow many prior messages each model receives as context. Full history is best for continuity but expensive on token count. "Prompt only" makes every agent respond fresh to just the original question.

Pause between agentsDelay in seconds between each model's response. Useful for reading along in real time, or for rate-limited endpoints.

Skip voted-off modelsVoted-off models (0 torches) are excluded from group chat runs. Disable to include everyone regardless of torch count.

Cross-talkWhen on, each model receives the prior models' replies as context — they can react to each other. When off, every model only sees the original prompt, like a blind survey.

Episodes

Auto-advance after Score blockBy default, episodes pause at Score blocks so you can rate responses before continuing. Enable this to skip the pause and keep running — scores are deferred until after the episode completes.

Default block timeoutSeconds before an individual task block is marked failed if the model doesn't respond. Set higher for large models on slow machines.

Scoring

Manual score sheetWhen you rate a response you score it 1–5 on each criterion below. The average becomes the manual score for that session. Appears on the score sheet after each run.

Auto-score signals — each response is scored automatically 0–10

Total weight 100%

Leaderboard

Default sortHow engineers are ranked when you open the leaderboard. Score uses the combined auto-score average.

Default viewCard grid shows rich detail per model. Compact list fits more on screen for quick comparison.

Show voted-off modelsWhen off, models with 0 torches are hidden from the leaderboard entirely. Useful once the roster gets large.

Data

Roster

BackupExports roster, teams, stats & timeline. Use before big changes or a new season.

RestoreOverwrites current roster & stats with a backup file. Cannot be undone.

Export persona promptsClean roster + prompts only — no scores or sessions. Hand to an AI to review or rewrite, then import back.

Import persona promptsUpdates prompts only. Matches by ID, then employee ID, then model path. Scores untouched.

Full app

Full exportEverything — roster, teams, machines, episodes, settings, custom roles, and presets in one file.

Full restoreRestores a full export. Overwrites all data including machines and settings. Cannot be undone.

Danger zone

Reset all statsZeroes scores, sessions, timeline, and episode count. Sets everyone to Unhired. Roster config, teams, machines, and presets are kept.

Clear rosterRemoves all models from the roster. Machines, teams, settings, and presets are kept. Cannot be undone.

Wipe all dataDeletes everything — roster, teams, machines, scores, episodes, presets. App reloads blank. Double-confirmed. Cannot be undone.