WolfPack LM Studio

LM Studio Hub: https://lmstudio.ai/krynexlabs/wolfpack-lmstudio — install with lms get krynexlabs/wolfpack-lmstudio or search "WolfPack" in the plugin browser.

Bring your project's .wolf/ knowledge base into every LM Studio chat. Local-only — no cloud, no MCP. Built for people who run a local model (e.g. Qwen) in LM Studio and want it to remember the project the way OpenWolf makes Claude Code remember it.

What it does

Auto-injects a resume digest on every message (prompt preprocessor): the current quest from STATUS.md, the Do-Not-Repeat list from cerebrum.md, and recently fixed bugs from buglog.json — so the model continues with the project's hard-won context. Bounded (default 1500 chars) so it stays cheap on a small local model.
Gives the model tools (tools provider):
- recall(query) — keyword search across .wolf/ (cerebrum, memory, STATUS, buglog)
- read_wolf_file(name) — read the beginning of STATUS/cerebrum/memory/buglog/anatomy (default 3000 characters; for anything further in, the model is pointed to recall)
- remember(fact) — save a durable fact to .wolf/memory.md

Two modes

Pure local (default). Reads the .wolf/ directory directly — digest + a built-in keyword recall. No dependency, no network, no MCP. Everything stays on your machine.
OpenWolf MCP sync (optional toggle). Routes recall and the resume digest through the OpenWolf MCP server (openwolf mcp), giving you the real BM25 recall with citation ids, plus Claude's native Auto Memory — the same engine Claude Code uses. Requires the CLI on the machine. If it's unavailable, WolfPack silently falls back to pure-local, so the chat never breaks.

Same .wolf/ either way — it's one shared knowledge base, not a separate store.

Write isolation (read union, write per-agent)

A local model reads everything (canonical .wolf/ + its own notes) but writes only to its own area: .wolf/local/<agentId>/memory.md. It never modifies STATUS.md, cerebrum.md, memory.md, or buglog.json — the authoritative knowledge base maintained by openwolf / Claude Code. Give each model a distinct agentId (default qwen) and their notes stay separate. A stronger model (or you) can then read .wolf/local/*/ to see what each local model did, evaluate it, correct it, and promote what's worth keeping into the canonical files — without a local model ever overwriting them. Same "propose, never overwrite" rule the OpenWolf AI tasks follow.

Private notes stay private

Anything wrapped in <private>…</private> in a .wolf/ file is blanked before the plugin looks at it — it cannot reach the injected digest, recall results, or read_wolf_file. This is the same guarantee OPENWOLF.md gives, and the plugin reads the same files, so it keeps the same promise.

The rules:

Filtering happens in the two functions that read knowledge files (one for text, one for JSON), so a future feature that reads one is covered without having to remember this.

What it does NOT do (and why)

OpenWolf's Claude Code integration passively watches file reads/edits and maintains anatomy, a token ledger, and memory automatically. LM Studio is a chat, not a coding agent — there are no file-operation events to observe. So the automatic, zero-effort capture stays exclusive to the Claude Code / Codex hooks; here you get the context-injection + recall + memory core.

Install

From GitHub (this repo)

lms is the LM Studio CLI (bundled with LM Studio; run lms bootstrap once if lms isn't on your PATH). lms dev registers the plugin with your local LM Studio instance directly from this folder — no Hub account needed.

From the LM Studio Hub (once published)

Search WolfPack in LM Studio's plugin browser, or lms get krynexlabs/wolfpack-lmstudio.

Setup

Open the plugin's global settings in LM Studio and set Project root to the absolute path of a project that has a .wolf/ directory (created by openwolf-enhanced). It applies to every chat.
Chat with your local model — the digest is injected into the first message of each chat, and the model can call recall / remember.

Config

Global settings (set once, valid in every chat):

Setting	Default	Meaning
Project root	(empty → env `OPENWOLF_PROJECT_DIR` → working dir)	Which project's `.wolf/` to use
Local model id	`qwen`	Write namespace under `.wolf/local/<id>/`; an id made only of dots falls back to `local`
Sync via OpenWolf MCP	off	On = recall + resume go through `openwolf mcp` (real BM25 + citations + native Auto Memory)
openwolf command	`openwolf`	How to invoke the OpenWolf CLI for MCP sync (or an absolute path)

Per chat:

Setting	Default	Meaning
Project root for this chat	empty = global value	Use a different project in this one chat
Auto-inject resume digest	on	Prepend the digest to the first message
Max digest size (chars)	1500	Keep small for small local models; quest, Do-Not-Repeat and bugs share it
Max size of read_wolf_file (chars)	3000	How much one `read_wolf_file` call returns. Several larger reads overflowed an 8k-context model in testing, and it began naming files that do not exist

Upgrading from 0.2.0: project root, model id and the MCP settings moved from the per-chat settings to the global ones. Set them once in the plugin settings; old per-chat values are not read any more.

Develop & publish

Published by krynexlabs (owner in manifest.json).

AGPL-3.0. Companion to openwolf-enhanced.