An LM Studio plugin that gives local LLMs eight read-oriented git tools built on @lmstudio/sdk and simple-git. The plugin lets a model inspect any git repository under the chat's working directory — staged/unstaged status, diffs, commit history, branches, file content at any ref, tracked-file listings, and content search — and clone new repositories into that directory for further inspection. All tools shell out through simple-git's discrete-argument process spawning (never a shell), every path is validated to stay inside the working directory, and git_clone is the only tool that writes to the filesystem.

Every tool takes a repoName parameter — the directory name of a repository inside the chat's working directory. Paths that resolve outside the working directory are rejected before any git command runs.

Returns the current state of the working tree as JSON: staged, unstaged, untracked, conflicted, and renamed entries plus the current branch and tracking info.

Returns the unified diff of staged or unstaged changes, optionally scoped to a single file. Returns (no changes) when the diff is empty.

Returns recent commits as JSON, in simple-git's log.all shape (hash, date, message, author, refs).

Returns local or remote branch info plus the repository root path.

Returns { current, branches, repoRoot } as JSON.

Returns the contents of a commit or a file at a specific ref, with line-based pagination. When a file is supplied, the tool first checks blob size with git cat-file -s and refuses to load anything larger than 10 MB. Post-load output is also capped at 10 MB.

Returns { content, total_lines, from_line, to_line } as JSON.

Lists tracked files as a recursive flat array, optionally narrowed by a git pathspec, with offset/limit pagination. Output capped at 10 MB; narrow with a pathspec when listing very large repos.

Returns { files, total_count, from, to } as JSON.

Runs git grep over tracked files and returns matches as JSON, with offset/limit pagination. Raw output is capped at 10 MB; narrow with a pathspec or a more specific pattern when searching very large repos.

Returns { matches, total_count, from, to } as JSON. Each match carries path, line, and text parsed from git grep --no-color -n -z -I --full-name.

Clones a remote repository into the chat's working directory. The destination directory name defaults to the URL's basename (with .git stripped) and is validated to stay inside the working directory before the clone runs. This is the only tool that writes to the filesystem.

Returns a confirmation string of the form Cloned "<repoUrl>" to "<repoName>".

Install the plugin through LM Studio's plugin browser. Once enabled, the eight tools become available to any model that supports tool calls.

Requires Node.js ≥ 22.

git clone https://github.com/packern/git-tools.git
cd git-tools
npm install
npm run dev    # runs `lms dev`

npm run push publishes the plugin to the LM Studio Hub (lms push).

The plugin is intentionally narrow. Read-only by default, single write tool, no shell:

• Working-directory containment — every repoName (and the git_clone destination) is resolved through resolveAndValidateRepoPath and rejected if it escapes controller.getWorkingDirectory(). This is the only error that throws rather than being returned as a string — the harness surfaces it to the user.
• Read-only by default — git_status, git_diff, git_log, git_branch, git_show, git_ls_files, and git_grep invoke only read subcommands. No commit, push, branch-creation, merge, checkout, or reset operations are exposed.
• Single write tool — git_clone is the only tool that touches the filesystem; its destination is path-validated.
• No shell — all git invocations go through simple-git, which spawns processes with discrete argument arrays. There is no shell-interpreted string anywhere in the plugin.
• No flag injection — git_show, git_ls_files, and git_grep route their string parameters through the shared gitSafeString zod schema, which rejects values starting with - so they cannot be reinterpreted as git flags.
• Blob and output size caps — git_show checks cat-file -s before loading file blobs and caps post-load output at 10 MB. git_ls_files and git_grep cap raw output at 10 MB and ask the model to narrow with a pathspec when the cap is hit.
• Recoverable errors are returned, not thrown — every tool wraps its simple-git calls in try/catch and routes caught errors through formatGitError, per the SDK guidance to return recoverable errors as strings.

The plugin has no user-facing settings — every constraint (size caps, pagination ceilings, max ref/path lengths) is fixed in source. Repository inspection happens entirely against repos under the chat's working directory.

With the plugin enabled you can ask the assistant to inspect any repository inside the chat's working directory ("what changed since last commit?", "show me the README at v1.2.0", "find every TODO in this repo"), or ask it to clone something new for inspection ("clone https://github.com/foo/bar so I can ask questions about it"). The assistant will pick the appropriate tool — typically git_clone first if the repo isn't present, then git_ls_files/git_grep to locate things and git_show/git_log/git_diff to read them.

• npm run build — compile TypeScript (src/ → dist/).
• npm run dev — run the plugin in LM Studio dev mode (lms dev).
• npm run push — publish to the LM Studio Hub (lms push).
• npm run lint / npm run lint:fix — ESLint on src/**/*.ts.
• npm run format / npm run format:check — Prettier.
• npm run knip — dead-code / unused-export check.

No test suite is configured.

MIT — see LICENSE.