// the take
Wanwu is an enterprise AI agent platform from China Unicom, targeting Forward Deployed Engineers who need to ship RAG, workflow, and agent-based applications to corporate clients. It packages the standard Dify-like feature set — visual workflow editor, knowledge bases, MCP integration, multi-tenant auth — but adds a GUI agent (screen interaction via sandboxed Docker) and an Ontology Agent for structured business data reasoning. It's aimed squarely at Chinese enterprise deployments, with explicit Xinchuang compliance for domestic CPUs and databases.
The GraphRAG integration is a genuine differentiator — their MultiHop-RAG benchmark numbers show a real F1 improvement over both Dify and LightRAG, not just a checkbox claim. The GUI agent with Docker sandbox isolation is a practical answer to the legacy-system problem that most enterprise AI platforms ignore entirely. Apache 2.0 with no commercial restrictions is the right call for an enterprise tool that needs to be embedded in client systems. The multi-tenant architecture with cost isolation is properly designed, not bolted on afterward.
The codebase is a microservice sprawl — at least 12 separate Go services (bff, agent, app, assistant, channel, iam, knowledge, mcp, model, operate, rag, tidb-setup) plus a Python callback service, which means the operational burden for a self-hosted deployment is substantial and the Docker Compose setup is the only real path. Documentation lives in Chinese markdown files bundled inside the repo under configs/microservice/bff-service/static/manual, which is an unusual and fragile choice. The workflow module lives in a separate repo (wanwu-workflow) that changed in v0.2.0, making version pinning and reproducible builds a real headache. The GUI client is a separate Baidu Netdisk download, not open source, which breaks the 'open technology' narrative for that particular feature.
We read the README, the directory tree and a key source file in July 2026. Nobody pays to be listed. How we judge →