---
url: /concepts/ai-and-mcp.md
description: >-
  How an AI agent such as Claude reads and changes your tasks through MCP, and
  what it can and cannot do.
---

# Hule and AI

An AI agent such as Claude connects to Hule through **MCP**
([Model Context Protocol](https://modelcontextprotocol.io)) and works with your tasks the way you
would: it finds, creates and updates them, signed in as you.

> **An agent acts as you: it sees only what you can see, and changes only what you could change — through a smaller set of actions than the app gives you.**

You bring the agent you already use; Hule has no AI model of its own inside the app.

## Agent requests: made as you, under your access

```mermaid
flowchart LR
  YOU["You"]
  AG["AI agent, for example Claude"]
  MCP["Hule MCP server"]
  CON["Connection you approved"]
  WS["Your workspaces and tasks"]

  YOU -- "ask in plain words" --> AG
  AG -- "calls tools on" --> MCP
  MCP -- "acts through" --> CON
  CON -- "opens only what you can open in" --> WS
  YOU -- "approves or disconnects" --> CON
```

Every tool the agent calls turns into an ordinary request to Hule, made as you. The same
[sharing rules](/concepts/sharing) and [permission sets](/reference/permissions-and-access)
that decide what you see decide what the agent sees.

## Agent connection: approve once, disconnect any time

Hule's MCP server lives at this address:

```text
https://mcp.hule-do.com/mcp
```

You add it to your agent once, and the browser opens Hule's consent screen. It says what you are granting: the app will be able to read and change
everything you can. You press **Allow access**, and the agent is connected. You never paste a
password or a key into the agent. Setup for Claude.ai, Claude Desktop and Claude Code is in
[Connect an AI agent](/guides/connect-an-ai-agent).

::: info In Claude.ai: a custom connector
Claude.ai adds the address as a custom connector; the guide shows where.
:::

The connection then appears in **User settings** → **API tokens**
([Manage your account](/guides/manage-your-account)) next to your personal tokens, with an
**OAuth** badge. **Disconnect** there, and the agent loses access at once; you can connect it
again by approving it once more.

## Agent actions: what the agent can read and change

The agent works inside the tree you have already built — see [The Hule model](/concepts/hule-model):

| Area | What the agent can do |
|---|---|
| **Find tasks** | search, filter and count them |
| **Read a task** | with its [subtasks](/guides/break-down-with-subtasks) |
| **Change tasks** | [create, update](/guides/create-and-edit-tasks), move and delete them |
| **Descriptions** | write them with the same [rich formatting](/guides/write-a-task-description) as the editor |
| **Comments** | read comments and add new ones; edit or delete only your own ([Comment and mention people](/guides/comment-and-mention)) |
| **Relations** | read a task's relations, list relation types, and add or remove [relations](/concepts/relations) |
| **Tags** | put tags on a task or take them off; add, rename and remove a list's own tags; connect or disconnect a workspace tag set on a list ([Tag tasks](/guides/tag-tasks)) |
| **Time** | read time logs and log time on a task ([Track time on a task](/guides/track-time)) |
| **Notifications and files** | read and mark your [notifications](/guides/set-up-notifications), download and attach files |

It sees every workspace you belong to, so one request can cover your personal and work tasks
together — see [Personal and work tasks in one place](/concepts/personal-and-work).

## Structure and access: managed in the app

An agent's tools do not reach the structure or the access rules. It cannot:

* create, rename or delete workspaces, [folders](/concepts/hierarchy) or lists;
* create or change [views](/concepts/views), [dashboards](/guides/build-a-dashboard) included;
* set up [automations](/concepts/automations) or [recurring tasks](/guides/set-up-recurring-tasks);
* change [status templates](/concepts/statuses);
* [share](/guides/share-a-folder-or-list) anything,
  [invite members or assign permission sets](/guides/invite-members-and-grant-permissions).

You do those in the app, where you see the whole tree.

## Agent risks: what a wrong instruction can do

Because the agent has all your access, a wrong instruction reaches every task you can open.

::: danger Deletes can't be undone
A wrong instruction can permanently delete tasks, with their subtasks and comments. It can also
delete a list's own tag, which takes it off every task that carries it, or disconnect a tag set
and take its tags off the list's tasks. Hule cannot restore what the agent deletes.
:::

## Meeting notes to tasks: a worked example

You paste the notes from a planning call and write:

```text
Create a task in the Product list for each action item, assign the ones marked "me" to me,
due Friday, and link the task about the API to BT-0042 as "blocks".
```

**BT-0042** is a [Task ID](/reference/task-ids). The agent then:

1. Looks up the **Product** list.
2. Creates one task per action item.
3. Sets the assignee and the Friday due date.
4. Adds the **blocks** relation to **BT-0042**.

Each task appears in Hule at once for everyone who can see the list, and its **Activity**
([See what changed in a task](/guides/see-task-history)) shows the changes under your name.
Afterwards you can ask the agent:

```text
What is still open from Monday's call?
```

It searches the same tasks.

## AI guides and recipes

* Connect Claude: [Connect an AI agent](/guides/connect-an-ai-agent).
* Put it to daily use: [Create tasks from Claude](/recipes/tasks-from-claude).
* Understand the access the agent inherits: [Sharing along the hierarchy](/concepts/sharing).