// the take
TGO is a self-hosted customer service platform built as an AI-agent-first alternative to Intercom/Zendesk: LLM-driven agents handle chats across web, WeChat and other channels, backed by a RAG knowledge base, with handoff to human agents when the bot gets stuck. It's aimed at teams who want to run their own AI support stack rather than wire one together from an LLM API and a ticketing tool.
Human-AI handoff is a first-class concept (visitor tracking, session assignment, agent workspace) rather than a bolted-on escalation button. The MCP tool integration auto-generates interactive forms from OpenAPI schemas, which is a genuinely useful shortcut for wiring agents to internal APIs instead of hand-writing tool schemas. The workflow engine models agent logic as a typed DAG (LLM/API/condition/tool nodes), giving actual branching control instead of a single prompt chain. Channel coverage is unusually broad for an open-source project: web widget, native iOS, Flutter, WeChat Mini Program, with widget SDKs shipped separately from the core.
This isn't one app, it's ten-plus separately versioned services (six Python/FastAPI services, a TypeScript admin frontend, a Go device agent, native widget SDKs) glued together with docker-compose profiles — self-hosting means running a distributed system with its own ops surface, not deploying a binary. It hard-depends on WuKongIM for real-time messaging, an infra piece most teams won't already be running and now have to operate and trust. Stated minimums are 4 cores / 8GB RAM just to evaluate it locally, which is a steep bar before you know if the picks are good. The channel and deployment story (Gitee/Aliyun mirrors, WeChat/DingTalk/Feishu as primary integrations) is clearly built China-first; teams outside that ecosystem will be carrying integration surface they don't need.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →