AI Usage Policy
Effective October 28, 2026
| App | TreeCraft: Issue Hierarchies, a Jira Cloud app built on Atlassian Forge |
|---|---|
| Marketplace listing | Coming. TreeCraft is not on the Atlassian Marketplace yet, and its app key will be added here when it is. |
| Forge app ID | ari:cloud:ecosystem::app/2f99b935-aba0-4b23-8784-05691aaac4b9 |
| Vendor | Plugincraft Software inc., Québec, Canada |
The short version
- TreeCraft's only AI feature is a Rovo agent, TreeCraft hierarchy helper. The AI is Atlassian's Rovo and it runs on Atlassian's platform.
- The app has no model of its own and calls no AI service outside Atlassian.
- The agent acts as the person who asked, with that person's Jira permissions. Nothing it proposes is applied until the person agrees.
- Plugincraft receives none of the data the agent handles, so it cannot use it to train any model.
Scope
This policy covers the AI features of TreeCraft: Issue Hierarchies and the AI tools used to build it. It does not cover Rovo itself, which is Atlassian's product and is governed by Atlassian's terms and policies.
AI features of the app
The TreeCraft hierarchy helper agent works in Rovo chat on Jira sites that have Rovo. On a site without Rovo, nothing else in the app changes.
| Action | What it does | What it can change |
|---|---|---|
| Suggest rules | Proposes rules for a hierarchy from a description. It reads the rule kinds and settings, the hierarchy's rules (for people with Automate access), and the names of up to 50 projects, 50 boards and 50 link types that the person can see, and up to 100 of the site's fields. | Nothing. |
| Add rules | Adds the rules the person agreed to, at most 10 at a time. | Adds rules, the same way the Rules panel does. It never runs a write-back and changes no work item. |
| Summarize | Outlines a hierarchy, or the part under one work item: key, status, summary and assignee for each work item, with counts. At most 300 rows and 20,000 characters, and it says when it was cut. | Nothing. |
| Formula reference, Check formula | Gives the model the formula language, and checks a formula against it and the site's field names (up to 150 fields). | Nothing. |
| Check query | Checks a JQL query with Jira, as the person, or a tree query with the app's parser. | Nothing. |
| Save formula column | After the person agrees, saves a formula column. | One column, in the person's own library, never shared. |
Buttons in the app (Ask Rovo, Build with Rovo and Summarize with Rovo) open Rovo chat with a prompt for what the person is doing. The person can change the prompt before sending it, and nothing else happens until they send it and confirm what the agent proposes. The buttons show only on sites with Rovo.
Which model, and where data goes
The model is Atlassian's own, in Rovo. TreeCraft does not choose, host or train a model, and makes no call to an outside AI service. The agent's actions run as Forge functions on Atlassian. The app gives Rovo only what it needs to answer, as the person asking. The chat itself belongs to Rovo and is kept by Atlassian. The app keeps nothing from it, except a formula column the person agrees to save.
Work item details come from Jira as the person. A work item they cannot browse appears as "an issue you can't browse", with no key and no summary.
Human oversight and safeguards
- A person confirms every change. The agent proposes, and rules or a formula column are added only after the person agrees.
- The server checks again. Each action goes through the same checks as the app's pages: App access, the license gate, and the person's access level on the hierarchy. Added rules go through the Rules panel's own checks (Automate access, the rules capability and the rule validator). The model's rule text is treated as untrusted: a list of at most 10 rules, each of a known type with only that type's settings, and anything else is refused.
- Who is asking comes from the platform. The site and the person come from Forge's context for the call and never from the text the model writes. A call without them is refused.
- Text in work items is data. Summaries, folder names and memos can carry instructions aimed at a model. The agent is told to treat them as data. The only change an action can make is adding rules or saving a column, both confirmed by the person and checked by the server, so such text cannot run a write-back, change a work item or widen anyone's access.
- Answers can be wrong. AI output may be incorrect. Check what the agent proposes before you agree. The app also checks formulas with its own parser, and queries with Jira, before they are saved.
Optional MCP tools
The MCP tools are in preview. They are turned off by default, and a site admin turns them on.
A customer can choose to connect its own AI tools to TreeCraft through MCP tools offered with Atlassian's Forge MCP module. A site admin turns the tools on in Atlassian Administration and chooses which tools to expose, and each person consents before a client can act for them. A client then acts as that person, with that person's permissions.
- There are 16 tools. They read hierarchies, rules, views and settings, add memos and planning items, add and change rules, create and update hierarchies, and give formula and query help.
- No tool deletes anything or changes who may open a hierarchy.
- Suggest rules, Add rules and Summarize, which read work item details, are not offered as tools.
- The tools return ids, the app's own words and field names. They never return work item keys, summaries or field values.
- They share the REST API's switch and limits (120 calls a minute per person and 1,000 per site). An admin can turn the switch off for the site.
- Atlassian describes this module as in Preview, and use from outside clients as an Early Access Program feature, not for production.
Once an answer reaches a client the customer chose, the client and its provider handle it under the customer's own agreement with them. Plugincraft is not a party to it and cannot control it.
What we commit to
- The app will not send customer data to an AI service outside Atlassian. If that ever changes, we will update this policy and the Subprocessors page before the change takes effect.
- We will not use customer data to train or tune a model. Plugincraft does not receive it.
- The app's AI features suggest and a person decides. They make no decision about people on their own.
- We keep the agent's limits (rows, characters and counts above) and checks in place, and test them with every change.
AI in building the app
Plugincraft uses AI assistants to help write and review the app's code and this website. They work on the source code and on a development Jira site with sample data. No AI assistant is given customer data. Every change is also reviewed by a review agent that did not write it, and no change is merged without that review (see Security practices).
Questions and review
Questions about this policy: write to info@plugincraft.ca. We review this policy at least once a year.
Related documents
- Subprocessors
- Data Handling: Retention and Deletion
- Architecture Diagram and Data Flow (password protected: request access at info@plugincraft.ca)
- Security overview
- Privacy policy