Gremlin brings browser-native multi-agent coordination to the local machine

Gremlin is a browser-native multi-agent coordinator that runs locally and does not require a separate backend server, according to the project’s description. The tool is built with a TypeScript coordinator and a Svelte user interface, and it is designed to route messages between multiple agents while showing every exchange in the browser.

The project’s main pitch is control and visibility. Gremlin displays inter-agent messages in real time, lets a user click into any agent thread during a run, and allows human instructions to be injected at any point. The coordinator keeps agent state in memory and handles message routing, while the UI shows the activity monitor, file tree, code viewer, and settings panels.

To get started, users pick a model provider, supply a key if needed, choose a model, and run a task. Gremlin supports local and cloud providers including Ollama, LM Studio, WebLLM, OpenRouter, Groq, OpenAI, Anthropic, Google Gemini, Together, and custom OpenAI-compatible endpoints.

For local setups, the project points to Ollama and LM Studio. Ollama is described as local and free, while LM Studio can run a local server after a model download. Gremlin also supports WebLLM, which runs quantised models entirely in the browser using WebGPU, with no external server or API key required.

The application includes seven built-in modes: General, Engineering, Finance, Industrial, Medicine, Networking, and Prediction Markets. Each mode loads a different agent team, such as CEO, Researcher, Analyst, Critic, and Writer in General mode, or CTO, Frontend Dev, Backend Dev, Security Eng, QA Engineer, and others in Engineering mode.

The Engineering mode adds file system tools through the File System Access API. When that mode is active, an Open Folder button appears and gives agents access to read, write, and list files in a local project directory. According to the project, written files appear in a file tree in the left sidebar and can be opened in the code viewer.

Gremlin also includes web search and browser automation tools. DuckDuckGo is the default search provider, while Brave, Serper, Tavily, SearXNG, and Cloudflare-based search or crawling can be configured in settings. A sidecar server at 127.0.0.1:3131 enables agents to navigate pages, click elements, type into forms, evaluate JavaScript, and assert page conditions.

The project is intended to run from the local machine with npm run dev, which starts the Vite development server and provides a built-in CORS proxy. The build output is a single inlined HTML file, but the project says the full experience depends on the dev server because static hosting loses the proxy and sidecar features.

Gremlin’s message format is JSON-based, with agents returning analysis, routed messages, and a done flag. The coordinator parses those messages automatically, and plain text is accepted as a fallback final result. The project is released under PolyForm Noncommercial 1.0.0 and is intended for personal and noncommercial use.