AI导哈

GenKitKraft：LLM Agent 配置与运行平台

GenKitKraft 是一个可自行托管的 LLM 智能体平台，用于配置和运行基于 Google Genkit 的 Agent，支持多种大模型提供商、MCP 工具接入以及 OpenAI 兼容 API，适用于开发者快速构建和部署智能体服务。

  • by DEEJ4Y
  • Go
  • 本地部署
  • 5 星标
  • 更新于
在 GitHub 查看 DEEJ4Y/genkitkraft

使用说明

GenKitKraft

Self-hostable platform for configuring and running LLM agents. Built on Google Genkit (Go SDK). Configure providers, create agents with custom instructions, connect MCP tool servers, and expose everything through an OpenAI-compatible API.

Trial defaults: this deployment uses public default credentials (admin / demo-change-me) and a public default ENCRYPTION_KEY. Anyone who knows them can sign in to your app. After you deploy, update the three environment variables ENCRYPTION_KEY, AUTH_CREDENTIALS, and PUBLIC_API_KEY. Do this before you add a real LLM provider key. See the App Platform guide for the steps.

Features

  • [x] Any LLM provider: Google AI, OpenAI, Anthropic, Vertex AI, Bedrock, Azure, xAI, DeepSeek, Ollama
  • [x] Agent builder UI: Create agents with system prompts, model selection, temperature, tool config, and tool call limits
  • [x] Built-in tools: web_fetch fetches any URL and returns it as Markdown — no external setup required
  • [x] MCP tool support: Connect tools via MCP servers (SSE and Streamable HTTP transports)
  • [x] MCP server: All management APIs exposed as MCP tools — build agents on top of GenKitKraft with any MCP client
  • [x] OpenAI-compatible API: /v1/chat/completions with streaming support, works with any OpenAI client
  • [x] Embeddable chat widget: Design a website chat widget in the agent's Widget tab (or the public widget builder) and connect it through your own backend: integration guide
  • [x] Single binary: Frontend embedded in the Go binary; SQLite by default (zero external deps), PostgreSQL, MySQL, and MariaDB supported for multi-instance deployments
  • [ ] Smart tool selection: Three tool modes (manual, auto-search, hybrid) to avoid context pollution

MCP Server — Build Agents Programmatically

GenKitKraft exposes all management APIs as MCP tools. Connect any MCP client (Claude Desktop, Cursor, custom agents) to create providers, agents, prompts, and tools — no UI needed.

// Claude Desktop config (~/.config/Claude/claude_desktop_config.json)
{
  "mcpServers": {
    "genkitkraft": {
      "command": "npx",
      "args": ["mcp-remote", "http://localhost:8080/mcp"]
    }
  }
}

A built-in create-agent prompt guides the LLM through the full workflow. See the MCP Quickstart and Agent Creation Guide for details.

Docs

https://DEEJ4Y.github.io/genkitkraft/

Developer docs live under docs/:

API spec implementations and generated OpenAPI output are in spec/.

The chat widget builder UI shared by the dashboard and the docs website lives in widget-builder/.

License

MIT

相关 MCP