Most document chatbots are a vector search and a prompt. AgentForge is a full agent loop: route, retrieve, generate, call tools, finalize, with conversation memory in a SQLite checkpointer and tool errors fed back to the model.
Retrieval is hybrid, FAISS dense plus BM25 sparse merged with reciprocal rank fusion, one index per agent. Tools are built-ins, HTTP webhooks, or any MCP server over stdio or streamable HTTP, with tool output wrapped as untrusted data. Anthropic and OpenAI sit behind a circuit breaker with per-model cost tracking.
Chat streams over SSE with retrieval, token, tool call, and tool result events. A 19-case eval suite covers RAG accuracy, tool selection, refusal, and multi-step runs, and it runs without an API key.