# Create and Reuse Workflow Tools

> Package a visual workflow with a typed input/output contract so Chat Agents, Agent V2 apps, and other workflows can invoke it as a tool.

Source: https://doc.fastgpt.io/en/guide/build/tools/workflow-tools

A Workflow Tool packages a multi-step canvas flow—combining HTTP requests, knowledge retrieval, code execution, or model reasoning—behind a strict input and output schema so it can be shared and called across applications in your team.

## Use Cases and Prerequisites

* **Use cases**: Encapsulate multi-step business operations (such as querying an API, transforming data in a sandbox, and formatting a summary) into a single callable tool for Chat Agent V2 or a workflow [Tool Call](/en/guide/build/workflow/nodes/tool) node.
* **Permissions**: Tool creation and publishing permission in the active team, plus read/use access from the calling application.

![Configure custom inputs and parameters in a workflow tool](https://cdn.jsdelivr.net/gh/labring/fastgpt-img@main/optimized/current/tool-variables.png.webp)

## Step 1: Create a Workflow Tool and Declare Inputs

1. Go to **Studio → My Tools**, click the first **Create** card in the list, and select **Workflow Tool** (the upper-right toolbar provides **Batch Manage**, **+ Folder**, and **Import**).
2. The canvas initializes with two contract nodes: **Plugin start (`pluginInput`)** and **Custom Plugin output (`pluginOutput`)**, which contain the **Custom Inputs** and **Custom Outputs** sections.
3. Click **New** in the **Custom Inputs** section of the **Plugin start** node to declare each parameter:

| Setting         | Description                                                                                                                             |
| --------------- | --------------------------------------------------------------------------------------------------------------------------------------- |
| **Field key**   | Alphanumeric identifier (such as `orderId` or `queryDate`) used in the JSON call payload                                                |
| **Label**       | Human-readable label shown in the UI                                                                                                    |
| **Value type**  | `String`, `Number`, `Boolean`, `Object`, typed arrays, or `Select Dataset` (allowing the caller to pass target knowledge bases)         |
| **Description** | Clear instructions and format examples; language models read this description to extract and populate arguments during Function Calling |
| **Required**    | Enforces that callers or models must supply the parameter                                                                               |

## Step 2: Orchestrate Nodes and Bind Outputs

1. Add processing nodes (such as HTTP Request, Dataset Search, Code Run, or AI Chat) to the canvas and connect the execution handle from **Plugin start** to your first processing step.
2. Reference the declared input fields in your processing nodes using the variable picker.
3. Connect the final processing step to **Custom Plugin output (`pluginOutput`)**.
4. Click **New** in the **Custom Outputs** section of **Custom Plugin output**, define return field names (such as `resultSummary` or `statusCode`), and bind each field to the corresponding upstream node output. See [Call Application and Workflow Tool I/O Nodes](/en/guide/build/workflow/nodes/run_app) for node reference details.

## Step 3: Debug Independently and Publish

1. Click **Run** in the top-right corner, supply known test values in the input form, and run the flow.
2. Expand run details to verify that **Plugin start** values were received, intermediate nodes succeeded, and **Custom Plugin output** returned the expected fields.
3. Click **Save and Publish** in the top-right corner. Calling applications always execute the active **Published version** of the workflow tool.

## Step 4: Attach and Invoke in Applications

* **In Chat Agent / Chat Agent V2**: In the App's **Tools** section, click **+ Choose** and select the published workflow tool. If any input field is configured as a static caller setting (such as a fixed secret or Dataset scope), fill it in on the attached tool card.
* **Static call in a Workflow**: Add the workflow tool from **Add Node → Tools** as a standard step on the canvas and map upstream variables to its input ports.
* **Dynamic call via a Tool Call node**: Connect the workflow tool below a [Tool Call](/en/guide/build/workflow/nodes/tool) node using the blue diamond tool handle.

## Troubleshooting

| Symptom                                                | Check                                                                                                                            |
| ------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------- |
| Model omits arguments or passes invalid formats        | Refine the tool description and parameter descriptions in **Custom Inputs** with explicit format examples (such as `YYYY-MM-DD`) |
| Tool runs but the Agent receives an empty return value | Verify that internal node outputs are wired into **Custom Plugin output (`pluginOutput`)** and republish                         |
| Internal tool edits do not take effect in callers      | Click **Save and Publish** inside the workflow tool editor to update the published snapshot                                      |

Next: Compare tool types in [Choose Tools and Skills](/en/guide/build/tools/overview) or configure [HTTP Tool Sets](/en/guide/build/tools/http_tools).