From intent to workflow
Describe what you need in natural language to generate a workflow graph, node documentation, Python node code, and a backend entry point
OPEN SOURCE / GENERATIVE WORKFLOW
From ideas to
AI workflows that
evolve and adapt
Generate, run, and evolve AI workflows from natural language; FlowX turns one-off answers into processes you can inspect, reuse, and improve
PROMPT FOR YOUR AGENT
Read https://github.com/AIpRoBuilder/FlowX/blob/main/skills/flowx-workflows/SKILL.md and https://github.com/AIpRoBuilder/FlowX#setup; install the skill for this agent and install and verify FlowX in a suitable Python 3.10+ virtual environment; connect local MCP if needed, guide me through secure LLM configuration, and show me how to create a workflow; do not ask for secrets in chat
Open source · Local-first · Built for your agent
FLOWX ENGINE GENERATIVE WORKFLOW
INPUTNatural language
FLOWXGENERATEGraph + nodes
RUN & EVOLVERun and evolve
DESCRIBEGENERATERUNEVOLVE↻
BUILT FOR AGENTS / MADE TO EVOLVESCROLL TO EXPLORE
01 / WHAT YOU GET
More than an answer
Build the process
FlowX focuses on the execution layer behind your agent: turn requirements into clear, runnable workflows that improve with feedback
Audited and tested nodes
Each business node goes through generation, audit, testing, and bounded repair; nodes that still fail are not published
Evolve through conversation
Change an entire workflow or specific nodes as needed; unchanged parts are reused as the process grows with your requirements
Inspectable and reusable
Workflow graphs, code, and versioned artifacts are stored locally so you can inspect, reopen, and evolve them later
02 / HOW IT WORKS
From a few words
to a reusable process
FlowX connects intent to execution, generating explicit graphs and code instead of one-off results hidden in a conversation
THE FLOWX LOOPINTENT → WORKFLOW → EVOLUTION
✳
Describe
Tell your agent about your goal and the process you need
◇
Generate and verify
Build the graph, nodes, and entry point; nodes are audited, tested, and repaired
▷
Inspect and run
Inspect the artifacts, start the service, and run the workflow interactively
↻
Evolve and reuse
Refine the workflow or its nodes from feedback and reuse what you've built
03 / QUICKSTART
Start here
Build your first workflow
After installing FlowX, connect your agent with MCP or use the Python SDK directly; you'll need Python 3.10+ and a configured LLM provider
01 / MCP
Connect your agent
Register the local service in your MCP client; replace the paths with your virtual environment and configuration directory; on Windows, use the Scripts executable
mcp.json
{
"mcpServers": {
"flowx": {
"command": "/absolute/path/to/venv/bin/flowx-mcp",
"env": {
"FLOWX_CONFIG_ROOT": "/absolute/path/to/config"
}
}
}
}Set FLOWX_LLM_PROVIDER, FLOWX_LLM_MODEL, FLOWX_LLM_API_KEY, and FLOWX_DEFAULT_WORKSPACE in the configuration directory's .env; never commit secrets to a repository
02 / PYTHON SDK
Build a workflow in code
Once your LLM is configured, use the SDK to build a workflow from natural language and inspect its artifacts
build_workflow.py
from flowx_sdk import FlowXClient
with FlowXClient(workspace="./workflows") as flowx:
result = flowx.build_workflow(
"Build a workflow that summarizes a document.",
await_repair_input=False,
).raise_for_failure()
print("Workflow:", result.workflow_id)
print("Artifacts:", flowx.artifacts(result.workflow_id))Configure the model variables in the SDK's environment; building may execute node tests, so review the artifacts and code before running the service
04 / ONE CORE, THREE INTERFACES
Connect to FlowX
your way
One shared core, three interfaces; connect workflow generation to an agent client, a Python application, or another agent
01 / AGENT CLIENT
MCP
Connect an MCP-capable agent over local stdio to create, inspect, run, and evolve workflows; builds run in the background with queryable status
LOCAL STDIO02 / YOUR APPLICATION
Python SDK
Use the shared core directly from a Python agent or your own application backend to build, inspect, run, and evolve workflows
PYTHON 3.10+03 / AGENT TO AGENT
A2A
Let other agents delegate workflow building to FlowX through Agent2Agent; the adapter currently supports non-streaming workflow building
NON-STREAMING✦
Your interface, your choice — FlowX is model- and UI-agnostic: your frontend renders interactions and events, while FlowX provides the workflow layer behind them
Explore the architecture ↗05 / GET STARTED
Let your agent
start building
Start locally and connect your existing agent or application; see the project documentation for installation, configuration, and interface details
Read the setup guide01
Prepare your environment
You'll need Python 3.10+ and an OpenAI-compatible model provider
02
Install FlowX
One installation includes the shared core, MCP, Python SDK, and A2A adapter
03
Configure and connect
Configure your model, API key, and workspace, then choose an interface
04
Build and evolve
Describe your requirements, inspect the build, and keep improving your workflow
Is FlowX a visual workflow editor?
No; FlowX is a generative workflow layer and does not include a user interface; your agent or application handles interaction and presentation, while FlowX generates, runs, and evolves the process behind it
Where are generated workflows stored?
Workflow artifacts are stored in your configured local workspace, including the workflow graph, node code, and backend entry point; you can inspect and evolve them later
What happens when a node fails its test?
Node building includes audit, testing, and bounded repair; you can provide repair guidance when needed, and nodes that still fail are not published
How can I connect to FlowX?
FlowX supports local stdio MCP, a Python SDK, and an A2A adapter; the A2A adapter currently supports non-streaming workflow building
LET'S BUILD / TOGETHER
Build what comes
next for your agent
Explore FlowX or join us in building generative workflows
Emailpeterxcx@gmail.com
