Gaunt Sloth

Gaunt Sloth (gth) is a command-line AI assistant for code review, PR analysis, Q&A, and interactive coding sessions. It runs on your machine against whatever model you point it at, with every prompt in plain markdown and no vendor lock-in.

Based on LangChain.js.

Documentation · Quickstart · Official Site · NPM

Why?

Gaunt Sloth is small, extendable, cross-platform, and can itself be a dependency in your project. It started as a code-review tool that fed PR and Jira contents to an LLM; we kept finding more uses (spinning it up inside an MCP project to simulate cases, for instance), so it grew into a general configuration-driven CLI.

• Minimum dependencies — CommanderJS plus LangChain/LangGraph.
• Extensibility — write a little JS to add a tool or provider, or connect an MCP server.
• No vendor lock-in — bring your own API keys.
• All prompts are editable markdown — you are in full control.
• Stateless code reviews — developers can't argue it into changing its mind.
• Runs on-premises with local models on your terms.

Get started

npm install -g gaunt-sloth
gth init anthropic        # scaffold .gsloth.config.json and pick a provider
export ANTHROPIC_API_KEY="sk-ant-..."
gth ask "what does this project do?" -f README.md

The Quickstart walks through this end to end, and Guides & Recipes covers real jobs — reviewing PRs in CI, coding against your own project rules, running a free local model, scripting it non-interactively.

What it does

A configuration-driven CLI you wire into your own workflows — you choose the model, provider, prompts, and tools, and gth orchestrates them.

• Code review & PRs — review a PR against its linked GitHub/Jira issue (gth pr 42 23), or a local diff before committing (git --no-pager diff | gth review). Stateless, non-zero on failure, so it drops into CI.
• Q&A, chat & coding — one-shot questions (gth ask "explain this" -f utils.js), interactive chat, and full code sessions with filesystem access.
• Model experimentation — swap models and providers from config, no code changes; a customizable middleware pipeline; every system prompt an editable markdown file.
• Controlled automation — define custom shell tools with parameter validation, connect to MCP servers (including remote OAuth), and talk to external agents over A2A.

Providers

Bring an API key from any one of: Anthropic, Google (Vertex AI & AI Studio), OpenAI, Groq, DeepSeek, xAI, OpenRouter, Hugging Face, or a local Ollama / LM Studio endpoint — plus any other LangChain.js-compatible provider via JavaScript config. Scaffold one with gth init <provider>; per-provider setup lives in Configuration → Providers.

Commands

gth and gsloth are interchangeable. The main verbs are init, ask, review, pr, chat, code, exec, eval, batch, workflow, models, config, and history. Running gth with no subcommand drops you into chat.

See docs/COMMANDS.md for every command, argument, and flag.

Workspace Packages

This repository is an NPM workspace monorepo. Most users only need the gaunt-sloth app; the libraries are published for embedding.

Package	Description
gaunt-sloth	Main CLI application (packages/app). Installs the gsloth/gth binaries. Most users only need this.
@gaunt-sloth/agent	Agent runtime (packages/agent): AG-UI server, A2A client, MCP utilities, filesystem and custom tools, middleware registry.
@gaunt-sloth/review	Review and Q&A modules (packages/review) with content sources (GitHub, Jira). Ships the gaunt-sloth-review binary for lightweight CI — no dependency on commander, MCP, or A2A.
@gaunt-sloth/core	Config system, agent infrastructure, LLM provider wrappers, and shared utilities (packages/core).
@gaunt-sloth/batch	Batch matrix runtime (packages/batch) behind gth batch and gth eval: runs a prompt-executable over a matrix of models and/or content-bound inputs. Ships the gth-batch binary and the EvalReporter contract reporters implement.
@gaunt-sloth/eval-reporter-junit	JUnit XML (Ant-JUnit flavour) reporter for gth eval (packages/eval-reporter-junit). Ships with the CLI as the built-in junit reporter.
@gaunt-sloth/eval-reporter-teamcity	Live TeamCity service-message reporter for gth eval (packages/eval-reporter-teamcity). Not bundled — install it and register it under reporters.

The two eval-reporter-* plugins are optional add-ons with their own release cadence, so they sit on a version line of their own; the other five are version-locked and released together.

What each package declares as its in-workspace dependencies: @gaunt-sloth/core depends on none of them; @gaunt-sloth/agent and @gaunt-sloth/review depend on @gaunt-sloth/core; @gaunt-sloth/batch depends on @gaunt-sloth/agent and @gaunt-sloth/core; and the gaunt-sloth app depends on @gaunt-sloth/core, @gaunt-sloth/agent, @gaunt-sloth/review, @gaunt-sloth/batch, and @gaunt-sloth/eval-reporter-junit. Both eval-reporter-* packages take @gaunt-sloth/batch as a peer dependency rather than a runtime one, and nothing in the workspace depends on @gaunt-sloth/eval-reporter-teamcity — it is published for consumers to install themselves.

Installation

Requires Node.js 24+ (see the engines field in package.json).

npm install -g gaunt-sloth

latest follows the 2.0 line. Every release is published to its channel tag first (@beta while 2.0 is in beta), and a version that has proven itself is then promoted to latest — so the bare name gives you the newest release we have blessed, and gaunt-sloth@beta the newest one we have built. Upgrading from the old gaunt-sloth-assistant package? See Upgrading from gaunt-sloth-assistant (1.x)? first.

Configuration

Gaunt Sloth runs from a directory tree containing a config file (.gsloth.config.json, .gsloth.config.jsonc, .gsloth.config.js, or .gsloth.config.mjs), in the project root or under .gsloth/.gsloth-settings/. Create one with gth init, or point at one with -c /path/to/config.json.

The full surface — providers, tools, MCP servers, content sources, prompts, and identity profiles — is documented in the Configuration guide.

Uninstall

npm uninstall -g gaunt-sloth
rm -r ~/.gsloth        # global config, if any
rm -r ./.gsloth*       # project configs, if any

Contributing

Contributions are welcome through GitHub Issues and pull requests. For contributor workflow, local setup, and testing expectations, see CONTRIBUTING.md; participation is covered by the Code of Conduct.

Filing a bug? Run /debug-dump in a chat/code session to attach a diagnostic archive to the issue — see docs/debug-dump.md.

License

MIT. See LICENSE.