Jira for AI agents
Voice and chat AI agents can read and write Jira through a fixed set of tools. The tools use the same connection as the call features, and they follow the same rules: tenant isolation, sealed credentials and an audit entry for every call.
The tools
| Tool | What it does | Permission |
|---|---|---|
jira_search_issues | Searches issues with JQL that is escaped before it is sent. | jira:read |
jira_get_issue | Reads one issue. | jira:read |
jira_check_sla | Reads the service level of a customer's requests. | jira:read |
atlassian_knowledge_search | Searches the Confluence and Jira Service Management knowledge base that is indexed for the company. | jira:read |
jira_create_issue | Creates an issue in the mapped project. | jira:write |
jira_add_comment | Adds a comment to an issue. | jira:write |
jira_update_issue | Changes an issue's summary, description or priority, or moves it along a workflow transition such as Done. | jira:write |
An agent has the tools only when its allow list names them. The allow list is part of the Jira mappings (ai_tools in PUT /integrations/jira/mappings: agent id to tool names), so saving it needs jira:write, and that is the permission the agent's calls run with. An agent without an entry has no Jira tool. Without a connection, no tool is offered. Created issues go to the project and issue type of the mappings.
Writes need a confirmation
A write does not happen on the first request. The tool first returns a preview. The write runs only when the same arguments are confirmed in a later turn of the conversation, and each confirmation works once. A dry run shows what would be written without writing it.
What the agent sees
Text from Jira, including issue descriptions and knowledge articles, reaches the model as data, not as instructions. Results are cut to a fixed size.
Knowledge base
The knowledge base holds Confluence pages and Jira Service Management articles as plain text. In the mappings (knowledge) you name the Confluence space keys and the service desk ids and switch the sync on. Every six hours the pages are read through the same guarded connection and added to the company's AI knowledge base as sources titled confluence: <title> or jsm: <title>. Articles that change are indexed again, and articles that are removed from Atlassian are dropped from the index. If Atlassian cannot be reached, the last index stays in use. Sync now on the Jira settings page runs the sync at once for the saved spaces and service desks, and the page shows when the last sync ran and what it did (added, updated, removed, unchanged, failed), or the error code when it did not complete. atlassian_knowledge_search searches only these synced articles; your own knowledge-base sources are not touched.
From an AI client (MCP)
An MCP client sees the same actions as named tools: searchJiraIssues, getJiraIssue (scope jira:read) and createJiraTicket, addJiraComment, updateJiraIssue (scope jira:write, granted by name). The tool descriptions tell the client's model to confirm every write with the user first, issue text is marked as untrusted content, and every call is audited like a voice agent's.
Audit
Every tool call is written to the audit log with its outcome and, for writes, the issue key. The text of the request and the text of the answer are not written there.
Availability
The tools are available to the voice agent. Each agent's allow list and the knowledge sync settings are managed on the Jira settings page (Jira for AI agents: tool checkboxes per agent, Confluence space keys, service desk ids and the sync switch), or through the API.