Flow

AI agents for Frappe.

---

Flow puts an AI assistant inside your Frappe site. It knows your DocTypes, respects your permissions, and can operate your site in plain English — creating records, sending emails, building DocTypes, creating agents and tools, and setting up RAG chatbots.

Set up a support agent:

"Create an agent that resolves customer support issues. It should use the Support Ticket doctype and these files as its knowledge: [attach files]"

Flow sets up the RAG pipeline, embeds and indexes the files, and keeps the DocType source in sync — automatically.

Put reporting on autopilot:

"Every weekday at 9am, email me a summary of yesterday's new leads with their contact details and source."

Do anything a user can do:

"Find all overdue sales orders from last month and mark them as closed."

---

Quick Access

You can open the global AI chat interface from anywhere in the Frappe Desk by pressing:

Cmd + I (or Ctrl + I on Windows)

---

Framework

Flow also ships as a Python framework for building agents in code.

DocTypes

DocType	Purpose
Flow Provider	Provider-level credentials and endpoint settings
Flow Model	Model configuration — set a model_id like anthropic/claude-sonnet-4-6
Flow Tool	Reusable tools, defined as a module path or inline script
Flow Agent	Instructions, model, and tools — configurable from the Desk or in code
Flow Trigger	Run an agent on a DocType event or a schedule
Flow Knowledge Base / Flow Knowledge Source	Knowledge sources (files, URLs, DocTypes) for retrieval-augmented answers
Flow Session / Flow Run	Persisted conversations and execution history

Define a tool

import frappe
from flow import tool

@tool
def get_open_todos(allocated_to: str) -> list[dict]:
	"""List open ToDo items for a user."""
	return frappe.get_list(
		"ToDo",
		filters={"allocated_to": allocated_to, "status": "Open"},
		fields=["name", "description", "status"],
	)

Run an agent

from flow import Agent

agent = Agent(
	model="anthropic/claude-sonnet-4-6",
	instructions="You help with Frappe tasks.",
	tools=[get_open_todos],
)

result = agent.run("List open ToDo items assigned to me")
print(result.output)

Persist a conversation

session = agent.new_session(title="Todo assistant")

session.chat("What tasks do I have open?")
session.chat("Close the ones due before last Friday.")

Every turn is saved as a Flow Run linked to the session — so conversation history, tool calls, and outputs are all in the database.

Built-in tools

Tool	What it does
find_doctypes	Search for DocTypes by name or module
describe	Get the fields, filters, and available actions for a DocType or record
read	Fetch records from any DocType
create	Create one or more records
update	Update fields across multiple records at once
delete	Delete records
run_action	Call a whitelisted method or workflow action on a record
execute	Run arbitrary Python in a sandboxed context

Tool calls can require human confirmation before executing, and every run is persisted for audit.

---

Installation

Install into an existing bench:

cd $PATH_TO_YOUR_BENCH
bench get-app flow
bench --site site-name install-app flow

Then add a Flow Provider with your API key and a Flow Model with a model ID such as anthropic/claude-sonnet-4-6, and open the Desk Flow panel to start. For local providers like Ollama or LM Studio, set Base URL on the provider or model.

License

GNU AGPLv3