What does Jira Service Management in Claude actually do?
A service desk lead usually arrives at this from the queue view. Agents ask the same five things all day. Which requests have been sitting longest, what did the customer last say, which organization does this reporter belong to, is there a log file attached, how many open requests does this customer already have. Every answer is another tab.
Jira Service Management in Claude answers those questions and nothing else. Elaichi's Jira Service Management connector carries 30 tools, and every one of them is a read. Nothing in it creates, edits, transitions or deletes a request.
The tools split into four jobs. list_all_jiraservicemanagement_service_desks and list_all_jiraservicemanagement_queues describe the desk. list_all_jiraservicemanagement_queue_issues and list_all_jiraservicemanagement_search_issues pull the work itself. get_single_jiraservicemanagement_issue_by_id and list_all_jiraservicemanagement_issue_comments open one request and its thread. list_all_jiraservicemanagement_customers, list_all_jiraservicemanagement_organizations and list_all_jiraservicemanagement_organization_issues answer who is asking, and jiraservicemanagement_issue_attachments_download fetches a file off a request.
Atlassian's own Jira Service Management Cloud REST API covers the same ground: service desks, queues, requests, comments and attachments (developer.atlassian.com, checked November 2026).
Claude reaches all of it over MCP, at one address. MCP (Model Context Protocol) is the standard way an AI assistant calls tools in other apps. The address is https://api.elaichi.ai/mcp, the same string for every organization. The OAuth grant behind a client's token, not the URL, decides which organization and which person that client reaches.
Say the rest plainly. Agents still work tickets in Jira Service Management. Transitions, replies to customers, approvals and resolutions happen in Atlassian, done by the person doing them. This connector shortens the lookup, not the work.
Whose Atlassian access does each call carry?
The account the connection was made with, never the person typing in Claude. Every call through an Elaichi connection fetches that one connection's credentials, whoever is calling.
That leaves two shapes, and picking one is the first real decision.
Per-agent connections: each agent adds their own connection and signs in with their own Atlassian account. A queue listing then returns what that agent can see in Atlassian, and nothing beyond it. Atlassian's OAuth 2.0 (3LO) documentation says an app acts on a user's behalf, and that the permissions of that user constrain the app regardless of the app's scopes (developer.atlassian.com, checked November 2026). The same per-person shape, across three developer apps at once, is laid out in Claude Code with Jira, GitHub and Slack.
One shared connection: one person connects, then shares it at the use level. Every call runs on that single account for everyone it reaches. It is quicker to set up, and it gives the whole desk one permission set. Sharing a connection needs Gold or Black.
Do not blur the two by accident. A stored toolbox runs its entries on the connection pinned in them, so sharing one toolbox with the team does not give each agent their own Atlassian identity. For per-person reach, each agent uses their own connections, or stamps a shared template and picks their own connection while stamping. The template mechanic is covered in sharing one setup with a team.
Who can download a ticket attachment?
The person who ran the tool, anyone they share the resulting file with, and anyone who can use the connection or the toolbox behind the call.
The mechanic matters here. When a connected tool returns a binary body, a PDF, a screenshot or a log bundle, Elaichi stores it as a file and returns a download link and the file's details instead of the bytes. The MCP text says Elaichi saved the file, with its name and size, and notes that opening it needs a browser signed in to the Elaichi account. The attachment never lands in the chat as content.
Three groups can open that file. Its owner, meaning the person who ran the tool rather than the owner of the connection behind it. Anyone that owner shares it with, at view level. And anyone holding use access or better on the connection it came through or the toolbox it ran in, checked live and vetoed by that person's own restrictions. Org Owners and Org Admins get no oversight read of other people's files.
A file is capped at 100 MiB and expires 7 days after it is created, with expiry enforced on every read. Only its owner can share it, revoke a share or delete it. A person's own files are listed in the console under Settings, then Files. The audit log records file.created, file.downloaded with how it was fetched, and file.link_issued when MCP handed back a link rather than the bytes.
If attachments are the part you would rather not have on this desk, block that one tool for the roles that do not need it. DocuSign with envelope changes blocked works the same rule in a different app.
How do you restrict the connector for one role or one person?
With rules that name connectors and tools, written on a role or on one member. Governance restrictions come with Gold and Black.
A block rule naming the Jira Service Management connector and jiraservicemanagement_issue_attachments_download leaves the other 29 reads working for that role. Allow rules are the other shape, and they carry a trap. An allow rule turns on an allowlist for its whole layer just by existing. Once a role carries one, every connector and tool its allow rules do not name is blocked for that role, across every app. An allow rule that names nothing blocks everything for its target. Inside a layer, a block always beats an allow.
Rules on a member's role and rules aimed at that member both apply: A rule aimed at a member can only narrow what their role allows, and never replaces or loosens a role rule.
Two limits matter here. In Elaichi, restriction targets are role or user only; the organization default is the absence of a rule, which means allow everything. So holding a whole team off a tool means writing the rule on their role or on each member. And a restriction names connectors and tools, never records. It cannot say "only the first-line queue" or "only this customer's organization". Queue and project scope stays on the Atlassian side, in the account behind the connection.
A restriction change takes effect within about two minutes, not on the next call.
A restricted tool is not invisible either. A restricted tool is left out of the tools Elaichi offers a client and cannot be called. Search still names it, flagged restricted, with no schema. The refusal tells the agent that an admin can lift it. Any member can file that access request from their client. Only a person in the Elaichi console can approve or deny one.
Client-side controls will not do this job for you. In Claude's tool list, connected tools are never listed one by one, however few there are. What a client sees is search_tools, read_tool and execute_tool, so a control that picks individual MCP tools cannot single out a Jira Service Management tool. The MCP specification treats each tool as a name with its own schema, which is why a per-tool control can only act on the names a server advertises (modelcontextprotocol.io, checked November 2026). Per-app and per-tool choices belong in Elaichi, as covered in restricting one tool versus a whole app.
What does the audit log show for a Jira Service Management call?
One row for each connected-tool call that reaches execution, naming the person, the tool, the connection and the client that asked.
A row records the acting user, the toolbox the call went through, the tool's advertised and catalog names, the connector, the connection the call actually reached, the tool's classification, how approval was given, the surface and the MCP client, the outcome, what the call touched, and a result count for list and search calls. In the console it reads as a sentence that opens with where the call ran, for example "Using Claude".
What a row leaves out is just as useful. There are no argument names, no argument count, and no argument values beyond the one path argument that names the object. A list or search call records the resource type and how many items came back, nothing about the items. A target never keeps a title for a person's record or for a message, and an email address is never recorded as an id or a title. So the row says that someone listed requests in a queue and how many came back. The ticket text stays in Atlassian.
One gap to plan around: a call a restriction withholds writes no row, because the refusal lands before anything runs. Restriction changes themselves are recorded, as created, updated and deleted events.
Elaichi keeps 90 days of audit history. For a longer trail, forward it to your own log service: export to your own Datadog comes with the Black plan. The shape of that export is in sending tool-call logs to Datadog, and what a row has to contain is in the audit log checklist.
Every member can read the audit log. With the audit permission, held by Org Owner, Org Admin and the free Auditor role, they read every event in the organization. Without it, they read their own events plus events on connections they own or can edit.
What about prompt injection in a customer's comment?
A request comment is text a stranger wrote, and Elaichi's MCP endpoint does not scan tool results for injection. That defense belongs to the client.
A read-only connector bounds one part of the problem. Nothing in the Jira Service Management connector can be talked into commenting, transitioning or deleting, because no such tool exists in it. The grant is the wider surface. A grant holding "Use your connected apps" runs every tier of tool in every connected app it reaches, so text read from one request can still aim at another app that agent has connected. What limits it is the set of toolboxes the grant was approved for and the restrictions on the role.
Two practical notes follow. A grant can be limited at consent to between one and 50 chosen toolboxes, re-checked on every call. And a list result is capped at 200 items, so a question about a whole backlog needs a filter in the asking rather than "show me everything".
Why not Atlassian's own MCP server?
Because sometimes you should. Atlassian publishes its own remote MCP server, and its repository says it spans Jira, Confluence, Jira Service Management, Bitbucket and Compass (github.com/atlassian/atlassian-mcp-server, checked November 2026). If Jira Service Management is the only app this desk will connect, and every agent already has an Atlassian account, that is one vendor fewer, and the honest answer is to use it.
The case for Elaichi starts at the second app. One address that Claude, ChatGPT, Cursor and Microsoft Copilot all sign into. One place where role rules live, so the attachment rule and the ticketing rule are written the same way. One audit log holding both calls. And one offboarding action: removing or suspending a member revokes their grants, and their clients fail on the next call. A team on Gold can also bring another remote MCP server in as its own connector and govern it under the same rules. The trade-off is laid out in one plane versus a server per app.
A rollout order that holds up
Follow it in this order, because each step narrows what the next one can get wrong.
- Connect one Jira Service Management account in Elaichi and read what it returns. Open Connections, add a connection for the Jira Service Management connector, and sign in with OAuth.
- Decide per-agent or shared. Per-agent keeps each person's Atlassian permissions. Shared gives the desk one permission set, on the owner's account.
- Give the desk a toolbox. Build one on the connection and share it at use, or share a template so each agent stamps it with their own connection.
- Write the restriction before anyone connects. A block rule on the agent role is the usual first one, naming the attachment download tool.
- Have each agent connect Claude once and approve their own grant, following the Claude setup.
- Read the audit log after a week, filtering to the connector, and see which tools the desk actually used.
If you are writing rules for the first time, start with who may do what, in which app. If the same team also lives in a ticketing tool outside Atlassian, the Zendesk setup is the nearest sibling to this one. The rest of the apps are in the connector catalog.