---
url: /agentic-ai/agent-datasets.md
---
# Extend an agent with datasets

An agent gets its instructions, skills, tools and working files from the datasets you connect to its input ports. You do not rebuild the agent: you register a dataset, give it the right **category**, and connect it in the *Workflow Designer*.

The category is what tells the agent how to use the dataset. The same folder is an instruction set with the category `Prompt` and plain working data with no category.

## Create the dataset

Registering a dataset for an agent is the normal dataset registration, with one extra step: choosing the category.

1. Prepare a folder with the layout the category expects (see [Categories](#categories)).
2. Go to **Datasets** and register a new dataset, uploading the folder. The steps are the ones of the [quickstart](../quickstart.md#_2-dataset-upload).
3. In the registration form, set the **category** of the dataset: `Prompt`, `Skill`, `Tool` or `Data`.
4. Save the dataset.

::: warning A dataset without category is data
If you leave the category empty, the agent treats the dataset as working data: its instructions, skills and tools are not loaded. This is the most common reason why a skill "does not load".
:::

## Categories

| Category | Folder layout | OpenCode Agent | ADK Agent |
| --- | --- | --- | --- |
| `Prompt` | `AGENTS.md`, `skills/`, and for OpenCode also `opencode.json`, `agents/`, `tools/`, `lib/`, `plugins/`, `docs/`, `init.d/` | The whole dataset becomes the agent's configuration. See [Prompt dataset](opencode-agent.md#prompt-dataset). | The root `AGENTS.md` is appended to the agent's instruction and `skills/` are imported. The rest is ignored. |
| `Skill` | `skills/NAME/SKILL.md`, one folder per skill | The skills are imported. | The skills are imported. |
| `Tool` | `tools/*.ts`, `lib/`, a root `package.json` | The tools are imported. | Skipped: ADK has no TypeScript tools. Use an [MCP server](adk-agent.md#connect-datasets) instead. |
| `Data` or none | Any files | Copied to `/workspace/data`. | Skipped. |

The plural forms (`Prompts`, `Skills`, `Tools`) are accepted. Any other category stops the agent at startup with `Unsupported category`.

::: warning Use only trusted Prompt datasets
On OpenCode a Prompt dataset is imported completely, including tools, plugins and permissions, and it can override the default permissions.
:::

## Connect the dataset to the agent

Both agents have six optional inputs:

| Inputs | Accept |
| --- | --- |
| `input-dataset-1` to `3` | URL datasets: MCP servers, and for ADK also A2A agent cards. |
| `input-dataset-4` to `6` | File datasets: the categories above. |

In the *Workflow Designer*, drag the dataset onto the canvas and connect it to one of the inputs `4` to `6` of the agent, then run the workflow. Datasets are read when the agent starts: after changing a dataset, restart the workflow.

You can connect more than one file dataset, for example a `Prompt` dataset with the instructions and a `Skill` dataset shared by several agents.

## Check it worked

Open the workflow logs with the show-logs button on the workflow page. Each dataset is either imported or reported with the reason it was skipped.

| Message | Cause |
| --- | --- |
| `Unsupported category ...` | The category is not Prompt, Skill, Tool, Data or empty. |
| `... category ... is not enabled for this runtime` | The agent does not support that category (for example `Tool` on ADK), or it was removed from `ALIDA_ENABLED_CATEGORIES`. |
| `... has no importable files` | The dataset is empty, or the files are not in the layout the category expects. |

**Next:** see the agent-specific details for [OpenCode](opencode-agent.md#connect-datasets) and [ADK](adk-agent.md#connect-datasets), or start from a ready-made Prompt dataset with [Service Writer](service-writer.md).
