Chatflow & Agentic Chat Settings

Beyond the canvas/prompt itself, every Agent application also has a set of application-level settings that shape its public identity, conversational behavior, and safety boundaries. This article covers the settings for the three modes—Chatflow, Agentic Chat, and Workflow—separately. Exact entry points and options follow the current product version.

Settings Overview for the Two Modes

Setting Category	Chatflow	Agentic Chat
Basic info (name/icon/description)	✓	✓
Opening message & suggested questions	✓	✓
Prompt (System Prompt)	Configured inside LLM/Agent nodes on the canvas	✓ Application-level
Model selection	Node-level (each node can differ)	✓ Application-level
Tool / Skill / knowledge base attachment	Attached inside the relevant nodes	✓ Application-level
Memory settings	Node-level (memory window)	Application-level
Guardrail	✓	✓
Publishing settings	✓	✓

The core difference: Chatflow's capability configuration is distributed across nodes (which node uses which model and which tools, each independently); Agentic Chat's capability configuration is centralized in the application settings (one prompt + one set of tools, with the model scheduling them autonomously).

Chatflow Settings

Entry: the settings icon in the top right of the canvas opens the "Settings" sidebar:

Setting	Description
Basic info	Application name, icon, description—the identity users see in the chat client and the Marketplace
Agent permissions	Edit permission (e.g. Workspace users) and usage permission (e.g. Workspace users only), with copy-link sharing supported (see Agent Permission Management)
Agent preset questions	Guiding questions shown to users, configurable separately in EN / Simplified / Traditional Chinese, lowering the barrier for the first question
Opening message	The welcome message at the start of a conversation; recommended to state the Agent's capability boundaries
Input variables	Custom fields defined by the Start node (see the Start node)
Memory window	How many turns of context to retain in multi-turn conversations, configured inside LLM/Agent nodes—use the smallest value that suffices to control usage
Enable cache / cache period	When enabled, cached results are reused for identical inputs, reducing repeated consumption; the cache period controls expiry
Simplified/Traditional Chinese conversion	When enabled, output is automatically converted between Simplified and Traditional per the user's language preference
Chat history search	When enabled, users can search their past conversations with this Agent
Input-box file upload	When enabled, the chat input box supports file uploads
Session variables	Session-level variables that persist state across turns for the flow to read and write
Environment variables	Application-level constants (e.g. API Key configuration) referenced by canvas nodes
Guardrail	The two-way input/output compliance interception switch; must be enabled for customer-facing Agents (see Guardrail)
Publish	Publish as a usable version; publishing to the Marketplace requires a security scan (Pro plan or above)

Agentic Chat Settings

Setting	Description
Basic info	Application name, icon, description
System Prompt	The application-level prompt: role definition, rules of behavior, output format, compliance red lines (for writing tips, see Prompt Writing)
Model selection	One application-level model; Mainstream / SOTA tiers, with SOTA costing more per unit
Tools	Attach built-in tools, third-party MCP (requires Premium), and OpenAPI tools; the model decides when to call them
Skill	Attach custom / official Skills as on-demand capability packs
Knowledge base	Attach knowledge bases so answers are grounded in your private documents
Human-in-the-loop	Allows the model to ask the user questions mid-execution and wait for answers
Enable to-do list	When enabled, the AI automatically identifies tasks in the conversation and breaks them down for execution (see Planning & Execution)
Spawn Subagent (Beta)	Allows the Agent to spawn independent Subagents via spawn_subagent to handle subtasks—suited to decomposable/parallelizable work; Subagents run in isolated contexts and return a summary when done
Agent permissions	Usage permission (e.g. anyone with the link) + copy link; can publish to the Agent Marketplace (see Agent Permission Management)
Simplified/Traditional Chinese conversion	When enabled, output is automatically converted between Simplified and Traditional per the user's language preference
Allow Fork	When enabled, other users can copy this Agent's configuration and adapt it for their own use
Guardrail	Same as Chatflow; Agentic Chat is more autonomous, so guardrails matter even more
Opening message & suggested questions	Same as Chatflow

Workflow Settings

Workflow has no chat client, so its settings are the most minimal:

Setting	Description
Agent permissions	Edit permission only (e.g. Workspace users)—Workflow has no chat client and no concept of usage permission; it is invoked by API/triggers
Enable cache / cache period	Same as Chatflow
Simplified/Traditional Chinese conversion	Same as Chatflow
Environment variables	Same as Chatflow; no session variables (Workflow is single-execution, with no session concept)

How Settings Relate to Plans

The available range of some settings is determined by the current Workspace's plan; options you can't select during configuration are usually not included in your plan:

• Third-party MCP tools: require Premium or above
• Official built-in Skills: require Pro or above; custom Skill counts vary by plan (3 / 20 / 100)
• Some models: models restricted by plan or region are not shown in the selector
• For the complete rules, see Configuration Time vs. Runtime

Related Reading

• Tip 1: Choosing Between Agentic Chat and Chatflow
• Debugging & Observability — how to test after your settings are in place
• Guardrail — a must-enable for customer-facing Agents