Tools, Workflows and Data Tables

These are LoopAgent's "capabilities to act" — once configured, the model can call them autonomously during a conversation to accomplish the task at hand.

Tools and MCP

• Once you bind plugins / MCP to a LoopAgent, the model can call the corresponding tools as needed.
• Credential security: sensitive information such as URLs, API keys, and auth headers is stored entirely on the platform side, execution also happens on the platform side, and the model never touches your secrets. Parameters named like key / token / secret are hidden from the model and injected by the platform at execution time.
• When there are few tools, the model sees and calls them directly; when there are many (more than 30), it switches to "lazy loading" mode to save tokens and control cost.

Key points

• A tool error does not interrupt the turn: when a tool call fails, the error is shown on the tool card, and the model relays it or retries and keeps serving.
• Can't see the tool-calling process: this is controlled by the "Show call process" switch under LoopAgent - Output (it must be on at both the plugin level and the agent level).

Troubleshooting when a tool can't be called

1. Whether the agent's tool master switch is on;
2. Whether the plugin is bound to this agent and enabled;
3. Whether that action is disabled at the plugin level or the agent level;
4. Whether the MCP inspection allowlist excludes that tool.

Workflows

A published workflow can become a tool that the model can trigger just like calling a tool; the input parameters of the workflow's START node become this tool's input parameters.

For a workflow to become a callable tool, three conditions must all be met (if any is missing, the tool does not appear, and no error is raised):

1. The agent's workflow master switch is on;
2. The workflow is associated with this agent;
3. The workflow is published (workflows that exist only as drafts are skipped).

• If you change a workflow draft without publishing, LoopAgent still runs the old published version.
• The workflow's "description" must clearly state its purpose, or the model won't retrieve it and won't call it.
• Credits for workflow execution are billed to the organization that owns the workflow.

Data tables and charts

Once you bind data tables to a LoopAgent, the model automatically gains two capabilities: querying tables in natural language, and turning query results into charts. The model never writes SQL.

• Natural-language table query: the model describes in a natural-language sentence what to query, and the platform turns it into a query and returns the results.
• Generate a chart: renders the results of the most recent query in this turn as a bar / line / pie chart and shows it directly to the user.
• Query first, then chart: before generating a chart, a successful query must have run in the same turn. If a user opens a new turn and simply says "draw a chart" without querying first, they are prompted to query once — this is by design, not a fault.

Table source: a LoopAgent's data tables are selected in its own data node. Select tables in the LoopAgent's data node; changing the data-table settings elsewhere on the agent has no effect on LoopAgent.