Flowcept Agent

The Flowcept Agent is an MCP-powered LLM interface for querying provenance data. It exposes a small set of tools (e.g., prompt_handler) that route natural-language questions to in-memory queries over captured task summaries.

Online-first design

Like Flowcept as a whole, the agent is designed to run while a workflow is still executing. In online mode, it consumes messages from the MQ (typically Redis) so it can respond to queries in near real time. This is the recommended setup for interactive RAG/MCP analysis during live runs.

Offline (file-based) queries

For simple tests or disconnected environments, the agent can also be initialized from a JSONL buffer file. In this mode, Flowcept writes messages to disk (dump_buffer), and the agent loads the file once at startup before serving queries.

This is a minimal offline example:

import json
from flowcept import Flowcept, flowcept_task
from flowcept.agents.flowcept_agent import FlowceptAgent

@flowcept_task
def sum_one(x):
    return x + 1

# Run a small workflow and dump the buffer to disk
with Flowcept(start_persistence=False, save_workflow=False, check_safe_stops=False) as f:
    sum_one(1)
    f.dump_buffer("flowcept_buffer.jsonl")

# Start the agent from the buffer file and query it
agent = FlowceptAgent(buffer_path="flowcept_buffer.jsonl")
# Or load a list of messages directly
# agent = FlowceptAgent(buffer_messages=msgs)
agent.start()
resp = agent.query("how many tasks?")
print(json.loads(resp))
agent.stop()

In the future, this page will include a full online example (live MQ + Redis) and deployment guidance.