Lm Studio Tier

lm-studio-tier

LM Studio's own skills, MCP servers and plugins are confirmed invisible to any API client — they fire only on the GUI's Send button or the native /api/v1/chat endpoint. A session bridged to LM Studio, or a script calling it directly, sees none of that. The only two things that actually reach /v1/chat/completions are the model and the preset (system prompt + sampling params) — everything below routes through those two levers only.

Lanes, models, presets

Assignment is measured, not guessed — see Documents/Local Model Roster.md. bulk-summarize's model is the one exception: provisional, chosen only to avoid the reasoning-budget failure qwen3.5-9b hit on the classify task, not itself benchmarked yet.

Presets live in ~/.lmstudio/config-presets/<lane>.preset.json — real LM Studio preset format (operation.fields keyed llm.prediction.*), so they're also visible from LM Studio's own GUI, not just from scripts.

Two ways to use this tier

Interactive session — Tools/local-tier/claude-local.ps1:

. .\claude-local.ps1
Start-LocalTier                    # idempotent, reuses server + model if already up
claude --bare --strict-mcp-config  # required, see below

Batch job — Tools/local-tier/lms-batch.ps1, for a list of items instead of a conversation:

pwsh .\lms-batch.ps1 -Preset classify -JobFile .\jobs\nodes.json -OutDir .\out\nodes

3-attempt cap per item, resumable (skips items already written), never marks an item done before its output file is confirmed on disk.

--bare --strict-mcp-config is required, not optional

Confirmed by reproduction (2026-08-20): running with the vault's full MCP config — roblox-studio's user_mouse_input tool schema in particular — breaks LM Studio's llama.cpp grammar compiler outright: Failed to initialize samplers: failed to parse grammar, a hard 400, not a slowdown. Both flags together, always.

Copy data into the isolated workspace — never --add-dir

claude-local.ps1 always cds into Tools/local-tier/workspace/ before starting the session. Work from inside that workspace; copy in whatever files the task needs rather than reaching out to the rest of the vault with --add-dir. The point of the isolated workspace is that nothing outside it can reintroduce the MCP-schema crash above or leak unrelated project context into a model with no judgment to filter it.

The honest latency ceiling

Local inference is slow, not fast — this is the tradeoff for zero token cost, not a bug to fix. Measured: qwen2.5-coder-7b-instruct at 7.90s for a short classification; gemma-3-12b at 47s for one image description; a 17GB MoE model at over 5 minutes just to load. Fine for a batch job left running in the background; wrong fit for anything wanting a quick turnaround or a back-and-forth conversation. Between runs, lms unload --all, and never retry a timed-out call without checking lms ps first — a killed call can end up competing with its own retry for the same compute.

What this tier cannot do

No tool use beyond what a model's own tool_use capability supports (check GET /api/v0/models, never assume from the model name — qwen2.5-coder-7b-instruct has none despite being a coder model). No MCP, no subagents, no vault memory, no multi-turn task tracking. If the job needs any of that, it stays on Claude Code proper.