English | 中文

LLM Space Web

An interactive learning workbench for building, running, tracing, and comparing AI Agents directly in the browser.

Try it online: https://kandian.site/llm-space-web/#/workbench

What you can do

• Start from nine editable Agent examples, including Weather Agent, Deep Research, General Agent, Translation, and Deep Wiki.
• Edit models, prompts, messages, variables, tools, MCP configuration, and ReAct settings in one workbench.
• Run supported Agents with streamed model output and safe browser/server tools.
• Inspect Run history and traces, compare executions, restore snapshots, and record evaluations.
• Run small browser-local Baseline/Candidate experiments with fixed text cases, deterministic checks, trace inspection, manual scoring, and JSON import/export.
• Learn with an Agentic AI assistant that explains the current surface, guides a weather-agent learning track, and performs only registered, confirmation-aware page actions.
• Keep guest Threads in the current browser and import or export them as JSON.

Hosted Alpha boundaries

This repository currently showcases the Web Hosted Alpha. It is not presented as a production multi-tenant SaaS.

• Guest Threads and virtual files are stored in the current browser.
• Free model usage is quota-limited; provider credentials remain on the server.
• Only allowlisted models and tools are available.
• Host Bash, host filesystem access, arbitrary DOM/JavaScript control, public remote MCP, login, BYOK, billing, and cross-device Thread sync are not available in the guest experience.

See the in-product Web quick start for the current user guide.

Local development

The repository uses mise as the task entry point and Bun as the package manager/runtime. Do not use npm, pnpm, or yarn for workspace dependencies.

mise run setup
mise run dev:guest-web

The guest Web app runs on http://localhost:15175/llm-space-web/.

Useful checks:

mise run check:changed
mise run build:guest-web
mise run pack:guest-api

Web architecture

apps/
  web/       # React/Vite workbench, docs, examples, and browser persistence
  cloud/     # Guest HTTP/SSE API, model quota, coach, and safe tool execution
packages/
  core/      # Browser-safe Agent/thread domain and server storage primitives
  ui/        # Shared Thread Playground and design system
  runtime/   # Model, tool, MCP, and Agent runtime implementations

The Web UI and guest API are built separately and served under the same origin in the Tencent Cloud deployment. Provider secrets never enter the browser bundle or local storage.

Project origin

This repository is a Web-focused fork of deer-flow/llm-space. The upstream project is a native Electrobun desktop application; this fork retains shared packages and desktop source for compatibility while the repository homepage, README, and hosted product focus on the browser workbench.

License

Released under the MIT License.