The agent platform that asks you before it writes the API wrapper

What it does AgentChat is a full-stack Python and Vue platform for deploying LLM agents through a web interface. It bundles dialogue, multi-agent task orchestration, RAG over vector databases, and tool calling into a single system backed by FastAPI. Users can chat with default agents or build custom ones that chain tools, query knowledge bases, and persist memory across sessions.

The interesting bit The standout feature is a human-in-the-loop workflow that conversationally generates MCP servers from OpenAPI specs, pausing at key decision nodes for human approval rather than blindly auto-generating code. The platform also implements a concrete three-tier memory architecture—short-term context capped at 3000 tokens, automatic historical summarization, and persistent long-term storage of user preferences—which is more deliberate than the usual “last N messages” approach.

Key highlights

• HITL-guided MCP Server generation from OpenAPI/Swagger specs with human checkpoints.
• Three-tier memory: 3000-token short-term window, auto-summarized history, and persistent long-term user profiles.
• Visual task-planning flowcharts for multi-agent collaboration.
• RAG support with Milvus, ChromaDB, and Elasticsearch backends.
• LangChain 1.0+ backend with a Vue 3/Element Plus frontend; Docker deployment supported.

Caveats

• The project is a heavy integration stack—expect to run MySQL, Redis, a vector database, and Elasticsearch alongside the Python backend.
• The README contains a broken API documentation link that points to a Google search placeholder, suggesting docs are still a work in progress.
• LangChain 1.0 introduced breaking changes in v2.2.0; older agent configurations may need migration.

Verdict Worth a look if you want a ready-made, visual agent workbench rather than building your own LangChain glue code. Skip it if you need a lightweight, single-binary assistant or are allergic to maintaining a full database fleet.