01 The Challenge
Coordinating Specialised LLM Agents Meant Hand-Writing Graph Code, Slow To Prototype And Hard To Hand Over To Non-Engineers.
Case Study · 03 / Multi-Agent Orchestration
Flowgentic
An Open-Source, MIT-Licensed Tool, Free For Anyone To Use, Fork And Self-Host, To Build And Coordinate Multi-Agent Teams: Wire Agents On A Canvas, Give Them Skills And Knowledge, And Ship Them Behind A Streaming API.
Product, Branding & Engineering ·
- 01 · Product
- Flowgentic
- 02 · Category
- Multi-Agent LLMOps
- 03 · License
- MIT · Open Source
- 04 · Released
- 2025
- 05 · Platforms
- Web · Self-Hosted · API
01 · By The Numbers
The Build,
By The Numbers
Counted Straight From The Codebase And Git History. No Vanity Metrics, No Projections.
5
Node Types
Root · Leader · Worker · Freelancer
2
Workflow Modes
Sequential · Hierarchical
3
LLM Providers
OpenAI · Anthropic · Ollama
45
REST Endpoints
Across 9 FastAPI Routers
02 · Overview
From Graph Code
To A Canvas
02 Our Approach
A Drag-And-Drop Canvas Where Agent Teams Are Wired Visually, Then Compiled Into LangGraph State Graphs, Sequential Or Hierarchical.
03 The Outcome
Teams Ship Multi-Agent Workflows With Tools, RAG And Human Approval Steps, And Expose Each One Through A Keyed, Streaming Public API.
03 · What We Built
Everything An
Agent Team Needs
Leaders Delegate, Workers Use Skills, Knowledge Comes From Your Documents, And Humans Can Step In Before Any Action.
Visual Team Builder
A React Flow Canvas With Snap-To-Grid Nodes; Drop An Edge On Empty Space And A New Worker Appears, Ready To Configure.
Skills
Built-In DuckDuckGo, Wikipedia, Yahoo Finance And Ask-Human, Plus Your Own HTTP Skills Defined In JSON.
Knowledge Base
Upload PDFs; A Celery Worker Embeds Them With fastembed And Stores Vectors In Qdrant For Retrieval.
Human In The Loop
Pause Any Agent Before It Acts. Approve, Reject Or Reply With Instructions Right In The Chat.
Public Streaming API
Per-Team API Keys And A Streaming Endpoint, With Threads Persisted Through A Postgres Checkpointer.
04 · Architecture
Canvas In,
State Graph Out
Team Config And Canvas Positions Live In Postgres; On Run, The Graph Builder Compiles Them Into A LangGraph StateGraph And Streams Tokens Back Over SSE.
- Sync · REST / GraphQL
- Async · Kafka Events
- Service Node
5 Node Types · 18 Migrations · MIT Licensed
Read The Architecture As Text
- Clients
- React SPA (Vite · React Flow), External Apps (Team API Keys)
- Edge
- FastAPI (9 Routers · JWT), Traefik (Reverse Proxy · TLS)
- Services
- Graph Builder (LangGraph StateGraph), Leader Nodes (Delegation), Worker Nodes (Tools · Skills), Celery Worker (PDF Embeddings), SSE Stream (Token Streaming)
- Data
- PostgreSQL (Teams · Checkpoints), Qdrant (Vector Store), Redis (Celery Broker), LLM Providers (OpenAI · Anthropic · Ollama)
05 · Stack
The Stack,
And The Code
A Typed React Canvas On A Strictly Typed Python Core, Every Model Provider Swappable Behind LangChain.
- 01 Frontend
- React 18
- TypeScript
- Vite 5
- Chakra UI
- React Flow
- 02 Backend
- FastAPI
- SQLModel
- Alembic
- Pydantic 2
- 03 AI
- LangGraph
- LangChain
- fastembed
- Qdrant
- 04 Data
- PostgreSQL
- Redis
- Celery
- 05 Infra
- Docker Compose
- Traefik
- Nginx
- GitHub Actions
- 06 Quality
- Ruff
- mypy strict
- pytest
- Biome
if i > 0: previous_member = members[i - 1] if previous_member.tools: graph.add_conditional_edges( previous_member.name, should_continue, create_tools_condition( previous_member.name, member.name, previous_member.tools ), ) else: graph.add_edge(previous_member.name, member.name) # Handle the final member's tools final_member = members[-1] if final_member.tools: graph.add_conditional_edges( final_member.name, should_continue, create_tools_condition(final_member.name, END, final_member.tools), ) else: graph.add_edge(final_member.name, END)$ docker compose up -d→ backend · frontend · celery · qdrant running06 · Open Source
Free For Anyone.
MIT Licensed.
Flowgentic Is Open Source Under The MIT License. Use It, Fork It, Self-Host It Or Build A Business On It. No Seats, No Lock-In.
View On GitHub ↗- $0Free Forever · No Seats, No Usage Fees
- MITUse Commercially, Modify, Redistribute
- 3LLM Providers, Incl. Local Models Via Ollama
- 1Command To Self-Host With Docker Compose
MIT License
A short and simple permissive license
Permissions
- Commercial use
- Modification
- Distribution
- Private use
Conditions
- License and copyright notice
Limitations
- Liability
- Warranty
$ git clone https://github.com/mry0tt4/flowgentic.git$ cd flowgentic# set SECRET_KEY, FIRST_SUPERUSER_PASSWORD, POSTGRES_PASSWORD in .env$ docker compose up -d# open http://localhost → Teams → New team
Have A Product
In Mind?
Tell Us What You’re Building. We’ll Reply Within One Business Day With A Scoped Plan.