// the find
fim-ai/fim-one
Open-source agent platform for Global × China enterprises — wire every system through one agent core. Self-hosted, any LLM.
FIM One is a self-hosted agent platform aimed at enterprises that straddle the Global/China software divide — it connects ERP, CRM, OA, and databases (including China-specific ones like DM, KingbaseES, and GBase that no other agent platform touches) through a single FastAPI/Next.js core. The three delivery modes (standalone assistant, embeddable copilot, central hub) cover the main enterprise deployment patterns without requiring separate installs. The target audience is a narrow but underserved one: IT teams in multinationals or Chinese enterprises that have to juggle both stacks simultaneously.
The China-stack database connectors are the real differentiator. Supporting DM, KingbaseES, GBase, and Highgo out of the box is not something you'll find in LangChain or n8n, and it reflects actual enterprise reality in China rather than a feature list padded for demos. The progressive disclosure meta-tools design is worth paying attention to — instead of handing the LLM the entire tool surface at once, it exposes a subset and fetches details on demand, which they claim cuts token usage 80%+. The Hook System running outside the LLM loop for human-in-the-loop approvals (posting an Approve/Reject card to a Feishu group and blocking until someone taps it) is architecturally correct — deterministic enforcement should never be inside the model's reasoning loop. The DAG planner with runtime decomposition, parallel asyncio execution, and up to three auto re-plan rounds is more production-grade than most agent frameworks that just chain LLM calls sequentially.
The license is 'Source Available,' not open source — multi-tenant SaaS, white-labeling, and competing agent platforms are explicitly prohibited. That matters if you're building a product on top of this rather than deploying it internally. The dev/prod database split (SQLite in dev, PostgreSQL in prod) is a parity gap that will cause surprises; JSON handling, constraint behavior, and query planner behavior differ enough that test environments won't catch real bugs. The RAG pipeline is tightly coupled to Jina AI for embeddings and LanceDB for the vector store — swapping either requires non-trivial refactoring, which is a risk if Jina pricing or reliability becomes an issue. Most of the messaging channels (Slack, Teams, WeCom, Email) are listed on the v0.9 roadmap rather than shipped; Feishu is the only working implementation today, so if your team lives in Slack the integration story is aspirational.