lmsgo

Delegate bulk I/O from Claude Code to a local LM Studio model. Single Go binary — no Python, no venv, no dependencies.

Core principle: Claude = thinking. LM Studio = I/O.

lmsgo ask     --question "..." [--glob PATTERN] [--max-tokens N] <file|dir> [...]
lmsgo write   --spec "..." --target <out> [--max-tokens N] [--dry-run] [ctx...]
lmsgo extract [-o output] [--last N] <session.jsonl>
lmsgo setup   [--model NAME] [--dry-run]
lmsgo version

Run lmsgo <subcommand> --help for the authoritative flag list. There are no short aliases (e.g. no -q for --question); Go's flag package does accept either single- or double-dash on the long name (-question ≡ --question).

Download and install LM Studio.

Open LM Studio → Discover tab → search for a model → download from Hugging Face. This must be done through LM Studio — do not download weights manually.

Tested models:

Any instruction-tuned model works. The system prompts were hardened against small-model quirks (preamble stripping, thinking-leak prevention) so most 4B–7B chat models should work out of the box.

The defaults LM Studio ships are tuned for chat, not the bulk-corpus pattern lmsgo uses. The values below are sized for Nemotron 3 Nano 4B — scale proportionally for larger models.

In LM Studio, click the loaded model and open its config panel.

Load tab → Context and Offload

Load tab → Advanced

Inference tab → Settings

Inference tab → Custom Fields

After changing these, reload the model so the new settings take effect.

Option 1 — lmstudio CLI daemon (recommended)

# Windows
irm https://lmstudio.ai/install.ps1 | iex

# Linux / Mac
curl -fsSL https://lmstudio.ai/install.sh | bash

lms daemon up

Option 2 — Desktop auto-start

Settings (Ctrl+,) → enable "Run the LLM server on login".

Docs: lmstudio.ai/docs/developer/core/headless

brew tap payfacto/tap
brew install payfacto/tap/lmsgo

Download the archive for your platform from the GitHub releases page, extract, and place the binary on your PATH.

go build -o lmsgo .
cp lmsgo ~/bin/

lmsgo setup

This detects your running LM Studio instance, lets you choose a model, writes the environment variables to a sourceable file, and appends the Claude Code routing snippet to ~/.claude/CLAUDE.md — all in one step.

lmsgo setup --dry-run   # preview without writing anything
lmsgo setup --model gemma-4-e2b-it   # skip the interactive model prompt

lmsgo ask --question "What files are in this project?" README.md

# Answer a question across files
lmsgo ask --question "Where are JWT tokens validated?" src/auth/

# Specific files
lmsgo ask --question "What env vars does this app require?" main.go config.go

# Directory with glob filter
lmsgo ask --glob "*.java" --question "Where is the DB pool configured?" src/

# Generate a file
lmsgo write --spec "Go table-driven tests for UserService" \
            --target internal/user/service_test.go \
            internal/user/service.go

# Preview without writing
lmsgo write --spec "Dockerfile for a Go app" --target Dockerfile --dry-run

# Documentation update after a session
lmsgo extract ~/.claude/projects/my-project/session.jsonl -o /tmp/chat.txt
lmsgo ask --question "What doc updates are needed? Give exact edits." \
          /tmp/chat.txt docs/architecture.md

lmsgo/
├── main.go                      # subcommand routing and usage
├── client.go                    # LM Studio HTTP client (complete + listModels)
├── ask.go                       # ask subcommand
├── write.go                     # write subcommand
├── extract.go                   # extract subcommand
├── setup.go                     # setup subcommand (embeds CLAUDE_MD_SNIPPET.md)
├── internal/version/version.go  # version injected via -ldflags
├── go.mod
└── CLAUDE_MD_SNIPPET.md

go build -o lmsgo.exe .

No external dependencies — stdlib only. The binary is self-contained.

1. Create mycommand.go with func runMyCommand(args []string).
2. Add a case "mycommand": runMyCommand(os.Args[2:]) in main.go.
3. Rebuild and copy the binary.