🧰 LM Studio Agent Toolkit

Turn a local model in LM Studio into a working coding agent.
It reads and edits your code, runs your tests, remembers your project, uses git, searches the web and reads your documents.
No API keys. No cloud. Nothing leaves your machine.

Quickstart · What's new · What it can do · Tips & tricks · Safety · Full reference

_{A real session in LM Studio. The model goes on to fix the regex and re-run the test to exit_code: 0. Every call was shown for approval before it ran.}

⚡ Quickstart

lms get cervezagua/agent-toolkit

For long chats, also get its companion: lms get cervezagua/agent-toolkit-loop.

Then, in LM Studio:

1. Load a tool-capable model — LM Studio marks these "Tool use". Give it 32k context or more.
2. Enable agent-toolkit — it appears as a chip under the message box.
3. Click the chip and set Project Folder to the folder you want the model to work in.
4. Ask for work: "explain this project", "fix the failing test", "what changed since last week?"

That's it. Keep LM Studio's tool call confirmation on, and approve calls as they come.

Needs Node.js 22+ and git.

git clone https://github.com/cervezagua/lmstudio-agent-toolkit.git
cd lmstudio-agent-toolkit
npm run setup

LM Studio fetches the dependencies itself. Re-run npm run setup after git pull to update.

[!IMPORTANT]
Upgrading from the five separate plugins? Remove coder-tools, memory-tools, git-tools, web-tools and ocr-tools from your chats — enabling them next to agent-toolkit gives the model two copies of every tool. Then set Project Folder once, and if you used web-tools or ocr-tools, switch on Web or Documents: they start off.

🆕 What's new

The fixes in 0.5 were found by evals: real tasks given to local models, checked by their outcome. All releases are on the releases page.

🧩 What your model can do

One plugin, six groups. Each group is a switch in the plugin's settings, because a long tool list makes small models choose worse.

Every tool and setting is in the full reference.

More plugins

For long chats: agent-toolkit-loop. Enable it next to agent-toolkit, and a chat keeps going past the model's context limit in the same window:

• It hides old tool output first, and summarizes the earlier conversation only when that isn't enough.
• It keeps your own words and a list of what was done, and the model can search the full transcript.
• /btw side questions are answered without derailing the task.

lms get cervezagua/agent-toolkit-loop

It takes over LM Studio's reply loop for the chats it's enabled in; switch it off and LM Studio's own loop is back.

Just one tool? Three of the Web tools also come as small plugins of their own:

They're the same code as agent-toolkit's, so don't enable one alongside agent-toolkit with Web on, or the model sees the tool twice.

A system prompt to go with it

Small models often know how to call a tool but not when to stop. agent-toolkit-prompt is a preset that tells the model to look before editing, keep a task list, check its work, and never repeat a call that already failed. It sets the system prompt and nothing else. Get it from the Hub.

💡 Tips & tricks

Get better results

• Write an AGENTS.md in your project: how to run the tests, which folders matter, what not to touch. It's loaded into every new chat automatically, so you stop repeating yourself. CLAUDE.md works too.
• Ask for plan mode on anything big. "Enter plan mode and plan how you'd add X": while planning, the file-changing tools disappear until the model presents its plan with exit_plan_mode.
• Point at the file when you know it. "The bug is in src/parse.ts" saves a model several searches and a lot of context.
• Tell it to prove its work. "…and run the tests to show it passes" turns a guess into a checked result. Or type /test.
• Use the commands.
  • /review checks your uncommitted changes without touching them.
  • /commit writes the message for you.
  • /init writes an AGENTS.md.
  • Add your own as .agent-toolkit/commands/<name>.md.
• Undo a whole message's work with /rewind. It lists your messages that changed files; /rewind 3 puts the files back to how they were before message 3.
• Give skills triggers:. A skill with triggers: [deploy] is loaded whenever your message says "deploy". Small models rarely load skills on their own.
• Read only the part of a page you need. "Fetch the pricing page, just the table" lets the model pass a CSS selector to fetch_url instead of reading the whole page into its context.
• Summarise a video instead of watching it. With yt-dlp installed: "What does this talk say about caching? ".
• Ask for research, not a single search. With Web and Research on: "Research what changed in Node 24, with sources". It reads several pages and answers with a numbered source after each claim. It works best with SearXNG.
• Follow a blog or a project's releases. "What's new on https://example.com/blog?" reads its feed, even from the home page.
• Ask for file names before contents. A model can grep for just the files that match, or just the counts, which costs a fraction of the context of full results.
• Made a mistake? Ask for undo_edit. It restores a file to how it was before the chat's last change to it.
• Start dev servers in the background. "Start the dev server in the background and check it came up" uses the task tools, so the chat isn't stuck waiting on a process that never exits.

Long chats

• Enable agent-toolkit-loop for chats that run long. It keeps the chat going past the context limit, and you can ask /btw side questions without derailing the task. If you'd rather start fresh, have the model call save_session_summary, then start a new chat that picks it up from the memory index.
• Tell it what to remember. "Remember that we deploy from the release branch" saves a memory, which every future chat sees in its memory index and can read in full.
• Let it search your earlier chats. Switch on Search Past Chats and ask "what was the command we used to deploy last week?". It's off by default, because those chats hold everything you've typed in LM Studio.
• Turn a solved problem into a skill. Switch on Let the Model Save Skills, and after a task say "save what you just did as a skill". You see the whole skill before it's written, and future chats can load it.

Choose and tune the model

• Bigger tool-use models plan multi-step work far better than small ones. Checked end to end with a 27B tool-use model.
• Switch off the groups you don't need. Fewer tools means better choices, especially on small models.
• Models with their safety training stripped out ("uncensored", "abliterated") are often worse at following tool formats and may drift between languages. If one starts answering in the wrong language, add "Always reply in English." to the chat's system prompt.
• Pick the models for OCR, the sub-agent and research from a dropdown in the Documents, Files and Web settings. It lists the models you have; after downloading a new one, turn the plugin off and on to refresh the list.

Make the tools faster

• Install ripgrep and searches get much faster. Results are identical either way.
• Install the GitHub CLI and run gh auth login, and the pull request and issue tools appear.
• Run SearXNG for web search. Without it, search falls back to DuckDuckGo, which often answers automated requests with a bot check. searxng/ sets one up in WSL on Windows; on macOS or Linux, any SearXNG with the JSON format enabled works.
• Reuse your skills. Skills are the standard <skill>/SKILL.md layout, so pointing Skills Directory at ~/.lmstudio/skills (LM Studio Bionic's folder) or ~/.claude/skills gives your local model the skills you already have.

When something looks wrong

Platform notes

• Windows: commands run in PowerShell, with real exit codes. The toolkit repairs a missing PATHEXT in LM Studio's plugin host, so node, python and git resolve as they do in your terminal.
• macOS: apps started from the Dock don't see Homebrew's folders, so the toolkit adds /opt/homebrew/bin and /usr/local/bin back itself. The browser defaults to Chrome.
• Linux: works out of the box; the browser defaults to Chrome.

🛡️ Safety

[!WARNING]
This plugin lets a model act on your computer. Approving a tool call is the same as running that command yourself.

For untrusted projects, switch off Allow Shell Commands or run LM Studio in a virtual machine. More in SECURITY.md.

⚙️ Why it behaves well with small models

• Mistakes come back as text, not failures. A bad path, ambiguous edit or git error returns Error: …, so the model can correct itself instead of the chat dying.
• Edits are exact replacements that must match once, so a model never rewrites a whole file to change one line. LF text matches CRLF files, and any edit can be previewed as a diff or undone.
• Nothing is silently lost. Long command output is written to a file and its path returned; long documents and search results page with offset. A command that floods output keeps its first and last 2 million characters, instead of crashing the plugin.
• Near misses point the way. A missing file comes back as "signup.js does not exist. Did you mean: src/signup.js?", and an edit whose text isn't found shows the closest lines in the file, so the next try can copy them.
• Cheap answers first. grep can return just file names or counts, read_document_text reads a PDF's text layer without a model, and fetched pages are cached for a few minutes.
• The same answers on every machine. grep returns identical lines with or without ripgrep.
• Skills are read whole. A description that wraps, is quoted across lines, or uses YAML's > or | blocks keeps the part that says when to use the skill.
• The model starts informed. Each chat opens with today's date, your AGENTS.md, the memory index and, in a git repo, the branch, uncommitted changes and recent commits.
• It knows when it's in the wrong place. Without a Project Folder the model is told it's in an empty scratch folder, instead of concluding your project is empty.
• Real exit codes. PowerShell's -Command collapses every failure to 1; run_command reports what actually happened.
• No shell in the middle of git. git tools run git and gh directly, reject refs that look like options, and never open an editor or a credential prompt.
• The browser is navigable by number. A snapshot lists [3] link "Docs" -> /docs, and the model clicks or types by that number.

🚧 Limitations

• LM Studio plugins can't rewrite earlier chat history. agent-toolkit-loop works around that by sending the model a compacted prompt, so the chat on screen stays complete, and LM Studio's context meter still counts all of it.
• A plugin can't add its own button to the chat, so the project folder is set in the plugin's settings rather than with a folder picker.
• All chats share one browser.
• Results depend on the model.

🤝 Contributing

Issues and pull requests are welcome. CONTRIBUTING.md covers the layout, the tests and how to add a tool. CI runs the tests on Windows, macOS and Linux. To see whether a change helps real use, run the evals it affects against a local model before and after.

📜 License

MIT