AIpRoBuilder/FlowX

Workflow compiler and evolution loop for AI agents

Summary

FlowX is a Python framework that compiles natural-language intent into executable, inspectable workflow graphs for AI agents. It exposes the same capability through three interfaces (local stdio MCP, a Python SDK, and an A2A server) so agents can generate, run, and update workflows over the same conversation rather than relying on an opaque plan. The generated artifacts (a workflow.json, per-node Python files, and a FastAPI entry point) are written to disk so they can be reloaded for refinement and reuse.

What it is useful for

This is for engineers building agents that need to turn a user request into a concrete process they can inspect, execute, and revise inside the chat. FlowX separates workflow generation from chat rendering: the agent keeps ownership of the conversation while FlowX handles compiling the workflow graph, generating per-node Python code, auditing it, and managing a per-workflow FastAPI backend.

The README lists several end-to-end examples built the same way. A ticket_triage workflow classifies and prioritizes support tickets and drafts replies. A stock_pressure workflow computes chip-distribution and market-maker pressure scores over user-supplied OHLCV data. A picture_to_svg workflow uploads a QR code image, strips personal info, and converts it to SVG. A policy_scrawler workflow uses Playwright to look up official websites of Chinese policy institutions. In each case, the prompt becomes a set of node implementations you can list, run, and refine.

How engineers can use it

The documented setup requires Python 3.10+. The README shows four installation paths (pip, Poetry, uv, conda) and a flowx-mcp console script.

python3.10 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
cp .env.example .env
flowx-mcp

You then fill in FLOWX_LLM_PROVIDER, FLOWX_LLM_MODEL, FLOWX_LLM_API_KEY, and FLOWX_DEFAULT_WORKSPACE, register the server in an MCP-capable client, and drive it from chat. The Python SDK skips MCP entirely:

from flowx_sdk import FlowXClient
client = FlowXClient(workspace="./workflows")
artifacts = client.create_workflow(
    "ticket_triage",
    "Build a workflow that categorizes and prioritizes support tickets.",
)

For agent-to-agent delegation, flowx-a2a --host 127.0.0.1 --port 8001 exposes a non-streaming A2A endpoint (message/send, tasks/get, tasks/cancel).

Documented limitations: the primary interactive path is local stdio MCP; the A2A adapter only exposes non-streaming generation; FlowX is a workflow layer rather than a hosted UI, so rendering per-step events from the generated backend is the host frontend's responsibility. External data sources (Playwright, market data feeds) require their own access and dependencies.