FIM One Enterprise AI Agent Platform
Wire every system you already run — global SaaS to the China stack — through one self-hosted agent core, with any LLM.
Evidence shows access controls (admin endpoints, JWT auth, org isolation) but no detailed least-privilege principle. User confirmation exists (FeishuGateHook approval cards) but only for sensitive tool calls, with limited implementation details. Data flow transparency insufficient; no clear explanation of data flow or user visibility. Sensitive data handling has encryption (Fernet) and signature verification but not comprehensive. Dependency security has Dependabot monitoring but no vulnerability scan results. External effects have approval hooks but not fully assessed. Rollback not mentioned. Source attribution has contributor list and license, but publisher identity unverified.
Self-consistency is good; README, config, and tests align. Dependency availability partial; some dependencies (e.g., dmPython) require manual install, affecting availability. Failure messages partially present (error recovery, reconnection) but not comprehensive.
Audience clear (global × China enterprises), scenarios diverse (Standalone, Copilot, Hub). Capability boundaries described (three deployment modes, tool types), but trigger precision (e.g., auto-routing) lacks detail. Environment fit good (multiple LLM providers, databases, deployment methods).
Information architecture clear (README, docs, architecture diagrams). Install notes detailed (Docker, local dev). Naming stability not explicitly guaranteed. Examples and FAQ present (README, docs). Known limitations partially mentioned (e.g., some features on v0.9 roadmap). License explicit (Source Available License). Versioning and changelog mentioned (Changelog link) but not in-repo. Maintenance responsibility has contributor guide and active development signs.
Output usability good (tool artifacts, Markdown conversion). Marginal value high (connects global and China systems). Cost-benefit mentioned (token savings) but no concrete data.
Claim traceability partial; feature claims in README not fully verified in code. Cross-source corroboration limited; relies mainly on single repo. Fact-inference separation insufficient; some claims may be inferred.
- Publisher identity unverified; treat with caution.
- License is Source Available, not OSI open source, restricting multi-tenant SaaS and competing products.
- Dependency dmPython requires manual install, may affect deployment.
- Security features (e.g., approval hooks) only cover sensitive tools, limited scope.
- Rollback mechanism not mentioned; needs confirmation.
What does this agent do, and when should you use it?
FIM One is an open-source agent platform designed for global enterprises operating in China, connecting their existing systems — from global SaaS to the China stack (Feishu, WeCom, DingTalk, DM, Kingbase) — into a single agent core. It offers three delivery modes: a standalone AI assistant, an embedded Copilot, and a centralized Hub for orchestration. Core capabilities include dynamic DAG planning, ReAct agents, a full RAG pipeline, vision-aware document processing, MCP connectors, a Hook system for approvals and audit, and IM channel support via an abstraction layer. Deployment is flexible, with Docker self-hosting and a managed cloud version.
FIM One connects enterprise systems through a unified agent core. It uses a ReAct loop and a dynamic DAG planner to decompose goals into dependency graphs executed concurrently. The platform includes an OpenAI-compatible LLM interface, a FastAPI backend, and a Next.js frontend. Connectors automatically register APIs, databases (PostgreSQL, MySQL, Oracle, SQL Server, DM, etc.), and MCP servers as agent tools with auth injection. The Hook system runs deterministic policies outside the LLM loop, such as human approval cards posted to Feishu. The RAG pipeline integrates Jina embeddings, LanceDB, hybrid retrieval, and reranking. Document processing converts PDFs, Office files, and images to Markdown with OCR via vision models.
- A global enterprise integrates SAP and CRM with Feishu or WeCom for unified querying and operations.
- An IT team self-hosts an AI gateway on-premises to connect to DM or Kingbase databases, ensuring data residency.
- Operations staff receive approval requests in a Feishu group for sensitive actions like high-risk commands or data exports, and approve/reject with one tap.
- Developers use the visual editor to build workflows with branches and parallel steps, exported as JSON for reuse.
- A support department embeds a Copilot in its existing ticketing system to draft replies with knowledge base citations.
- A multinational pushes task-completion notifications and summaries to employees via Slack, Teams, or Feishu.
What are this agent's strengths and limitations?
- Strong cross-border connectivity covering both global SaaS and Chinese local systems, including Feishu, WeCom, DingTalk, and national databases.
- Flexible architecture with multiple delivery modes and embedding options (iframe, widget, embed) for various deployment scenarios.
- Built-in Hook system for human approvals and audit logging, enhancing security and compliance.
- Progressive disclosure and dynamic DAG planning reduce token consumption significantly, by up to 80%.
- License is source-available, not OSI-approved, restricting multi-tenant SaaS and competing platforms.
- Deployment and configuration require technical expertise, with Docker, multiple components, and environment variables/API keys.
- Channel support is currently fully implemented only for Feishu; Slack and Teams are on the roadmap but not yet complete.
- The RAG pipeline depends on external services like Jina AI and LanceDB, requiring API keys and potentially incurring costs.
How do you install or deploy this agent?
Docker is recommended: clone the repo (git clone https://github.com/fim-ai/fim-one.git), set LLM_API_KEY in .env, then run docker compose up --build -d. Open http://localhost:3000 to create an admin account. For local development, you need Python 3.11+, Node.js 18+, pnpm, and uv; run uv sync --all-extras and cd frontend && pnpm install.
How do you use this agent?
After launching, access the web portal and create an organization and users. Configure the LLM provider in settings (e.g., OpenAI, DeepSeek, local Ollama). Add connectors via OpenAPI import, AI chat builder, or MCP servers. Create or import agents, choose the execution mode (ReAct or DAG), and deploy them as a standalone assistant in the portal, an embedded Copilot, or a Hub. Configure IM channels like Feishu and set up approval hooks. Build knowledge bases with RAG, uploading documents for indexing. For daily use, simply chat with an agent, which plans and executes tool calls automatically.