FIM One: An Agent Platform for Enterprises Spanning Global and China Software Stacks

The Problem FIM One Addresses

Enterprise organisations, particularly those operating in both international and Chinese markets, typically run separate software stacks that cannot communicate directly. The README describes the problem as a sprawl of ERP, CRM, OA, HR, finance systems, and messaging platforms across regions that do not integrate with each other.

Global platforms typically lack connectors for China-specific software such as Feishu, WeCom, DingTalk, DM (dameng), KingbaseES, GBase, and Highgo. FIM One is positioned as the single agent core that connects all of these, with the same agent handling both sides of the global-China divide.

The cloud version at cloud.fim.ai is available for early access without Docker or API configuration. The self-hosted version supports any LLM provider and runs on the organisation's own infrastructure.

Three Deployment Modes: Standalone, Copilot, and Hub

The README describes three ways to deploy FIM One. In Standalone mode, FIM One acts as a general-purpose AI assistant with search, code execution, and knowledge base access, accessible through the portal.

In Copilot mode, the agent is embedded into an existing host system's user interface through an iframe, widget, or embed. This mode lets teams add AI capabilities to an existing application without replacing it.

In Hub mode, FIM One acts as a central orchestration layer across all connected systems. Tools from every connected system are registered in the agent core and the agent can invoke them in a single planning step.

All three modes use the same agent core. The README's Mermaid diagram shows ERP, Database, Lark, CRM, OA, and custom APIs all connected bidirectionally to the Hub agent core.

Installing with Docker

Docker is the recommended installation method. After cloning the repository, copy the environment template and set your LLM API key:

git clone https://github.com/fim-ai/fim-one.git
cd fim-one

cp example.env .env
docker compose up --build -d

Once running, open http://localhost:3000 to create an admin account on first launch. The docker-compose.yml maps port 3000 to the Next.js frontend and port 8000 to the FastAPI backend. The compose file also configures PostgreSQL, Redis, and an optional Docker socket mount for the code sandbox.

For local development with hot reload:

uv sync --all-extras
cd frontend && pnpm install && cd ..
./start.sh dev

The start.sh script supports multiple modes. Running it without arguments starts Next.js and FastAPI at localhost:3000 and localhost:8000. The dev mode adds hot reload for Python and Next.js HMR.

Planning and Execution: DAG, ReAct, and the Hook System

FIM One offers two reasoning modes. The ReAct agent uses a structured reasoning-and-acting loop with automatic error recovery. The DAG planner decomposes goals into a dependency graph at runtime: independent steps execute in parallel via asyncio, and the system re-plans up to three rounds if a step fails.

Auto-routing classifies incoming queries and sends them to the appropriate mode. This is configurable via the `AUTO_ROUTING` environment variable.

The Hook System is described in the README as deterministic enforcement that runs outside the LLM loop. The first shipped hook is `FeishuGateHook`, which intercepts sensitive tool calls and posts a human approval card to a Feishu group before executing. The hook is described as extensible to audit logging, read-only-mode guards, and rate limiting in a future release.

Token budget management is handled by ContextGuard, a five-layer system that the README describes as part of the production-grade execution environment. Progressive disclosure meta-tools reduce the tool surface presented to the LLM to keep token usage tractable.

Database and System Connectivity

FIM One includes database connectors for PostgreSQL, MySQL, Oracle, and SQL Server, as well as four enterprise databases primarily found in Chinese enterprise deployments: DM (Dameng), KingbaseES, GBase, and Highgo. The README notes that schema introspection and AI-powered annotation are supported for connected databases, allowing the agent to understand table structure and query the data.

Actions (external system integrations) can be built three ways: by importing an OpenAPI specification, using the AI chat builder, or connecting an MCP server directly. Actions auto-register as agent tools with authentication injection.

The README describes a progressive disclosure design where meta-tools reduce the token usage required to interact with large tool sets by 80% or more across all tool types. This is a specific design choice for managing the cost of tool-heavy agents in environments with many connected systems, where naive tool listing would quickly exhaust the model's context window.

Limitations and Licence

The pyproject.toml lists the license as `LicenseRef-FIM-One`, a custom licence identifier. The GitHub metadata shows `NOASSERTION`, meaning GitHub could not identify a standard licence. This is not an OSI-approved open source licence, which affects whether the software can be used, redistributed, or modified under open source terms. Any commercial or redistribution use requires reviewing the actual licence text.

The README notes a known limitation with content guardrails: the default jailbreak-phrase detector runs before the LLM call, which means false positives could block legitimate requests. Output guardrails are optional, configurable via `FIM_GUARDRAILS_OUTPUT`.

The README also documents that extended thinking (chain-of-thought) is supported for OpenAI o-series, Gemini 2.5+, and Claude, but this requires models that support the feature.

The last push was on 2026-09-16. The project version in pyproject.toml is 0.8.12. It carries an "Alpha" development status classifier.

Comparison With Dify

Dify is another open-source LLM application and agent builder with a self-hosted option. It supports workflow construction, RAG pipelines, API publishing, and a visual editor. Dify is primarily built for teams that want to create LLM-powered applications and workflows without specific enterprise cross-border connectivity requirements.

The practical difference is connectivity scope. FIM One explicitly targets the combination of global SaaS and the China software stack. Its database connectors include DM, KingbaseES, GBase, and Highgo, which are absent from general-purpose platforms. The FeishuGateHook and native Feishu, WeCom, and DingTalk integration are specific to FIM One.

For a team that only uses globally common tools and databases, Dify's more mature tooling and documentation would likely be the easier starting point. For a team managing the global-China enterprise divide, FIM One's integration layer addresses needs that general agent platforms do not.

Editorial conclusion

FIM One is a specific fit for enterprises that need to wire together systems from both global SaaS and the China software ecosystem through a single agent interface. The DAG planner and Hook System address real production concerns around parallel execution and human approval gates. Before deploying, examine the LicenseRef-FIM-One license in pyproject.toml: it is not an OSI-approved open source license, and redistribution or commercial use terms require review.

Frequently asked questions

Official sources

1. fim-ai/fim-one on GitHub
2. Issues
3. Project website
4. README