LLM Space is a local-first desktop application for agent builders. It covers prototyping an agent, watching each model call and tool run inside the agent loop as it happens, replaying a run from history to find where it went wrong, and measuring how the agent performs across runs. Threads are stored as files on your own machine, and API keys stay local; the app collects a small amount of anonymous usage data with a documented opt-out.
Prompts, system messages, tools, and model settings are written and versioned in the app. It can generate prompts and tools, and turn a thread into a runnable LangGraph agent. Version 4 is the fourth major iteration of a project started in March 2023, and it is a sister project of DeerFlow, which is built and debugged with it.
The stack is TypeScript built with Bun, an Electrobun desktop shell, a React interface with Tailwind CSS and shadcn/ui, and Pi Agent Core as the agent framework. The repository is a Bun monorepo with core (domain types, clients, storage, generators), runtime (local runtime, models, tools, skills, MCP, plugins), ui (shared design system and thread playground), and the desktop app, plus an example plugin covering every extension type.
Two macOS editions are published for Apple Silicon and Intel: a smaller build that uses the system WebView, and a Performance build that embeds its own rendering engine for consistent rendering across macOS versions. Both share the same ~/.llm-space data directory and update in place, so switching editions keeps threads and settings.
Features
- Build: write and version prompts, system messages, tools, and model settings
- Trace: watch every model call and tool run inside the agent loop in real time
- Debug: replay a run from history and step through it
- Evaluate: measure agent performance across runs
- Manage: keep threads organised as files on your own machine
- Generate: have the app draft prompts and tools, or turn a thread into a runnable LangGraph agent
- Plugins: extensions covering skills, MCP servers, model providers, plugin tools, commands, thread storages, and settings schemas
- Two editions: a system-WebView build and a Performance build with an embedded renderer, sharing one data directory
