Linear vs Jira for AI coding agents: who hands out the work

Both trackers now hand issues to coding agents

The Linear vs Jira decision now includes how agents receive work. Both trackers have documented agent handoffs alongside their MCP servers. Those are different mechanisms: a handoff starts an agent session, while MCP gives a connected client tools to work with the tracker.

Linear launched Linear for Agents on May 20, 2025, making agents first-class workspace users. Its own Linear Agent is on by default. The coding sessions introduced on June 11, 2026 run Claude Code or Codex in a cloud environment and return a draft pull request.

Atlassian introduced Rovo Dev in Jira on April 2, 2026. Its July 15, 2026 announcement put Jira Coding Agent into paid Jira plans. Jira also accepts third-party agents through assignment and other triggers.

The distinction for a fleet is how the next issue reaches an idle agent. These tracker integrations start work from a delegation, assignment or trigger. A shared queue lets the agent ask for eligible work itself.

Linear separates human assignment from agent delegation

In Linear, the Assignee identifies a human responsible for the issue. An agent goes in the separate delegate field, which holds a single agent. Linear documents no restriction preventing a second agent from being mentioned on the same issue or opening its own session there.

A mention or delegation automatically creates an agent session. Linear says the agent must answer within 10 seconds or appear unresponsive, with up to 30 minutes before the session becomes stale. An agent can recover from that stale state.

Linear’s developer-preview agent APIs require the agent to acknowledge a delegation webhook within 5 seconds. That webhook carries issue details, comments and guidance in promptContext. The agent reports progress through activities such as thought, elicitation, action, response and error, plus a session-level plan. After a stop signal, it must take no further actions.

Linear’s agent integrations support delegation to Cursor, Codex, GitHub Copilot and Devin. Claude connects through Linear MCP, while Claude Code also connects through third-party apps or runs in Linear’s own coding sessions. Agents are not billable seats. The agent platform and MCP access are available on Free, Basic and Business. Linear’s own coding sessions use prepaid AI credits.

Linear MCP gives clients issue tools

The Linear MCP server URL is https://mcp.linear.app/mcp, using Streamable HTTP. Its read-only endpoint is https://mcp.linear.app/mcp/readonly.

Linear documents OAuth 2.1 with dynamic client registration, or an Authorization: Bearer header containing an API key. Linear removed SSE support in February 2026. An older configuration pointing at an SSE transport therefore needs to move to the documented HTTP endpoint.

Linear publishes no MCP tool list. Its docs describe tools for finding, creating and updating issues, projects and comments.

Jira starts agent work through fields and triggers

Atlassian’s July 15, 2026 announcement says Jira Coding Agent is built into every paid Jira plan. Its setup documentation names Standard or Premium, Rovo credits, and Bitbucket Cloud or GitHub Cloud as requirements. Those pages differ in plan wording, so check the setup requirements rather than assuming the announcement resolves eligibility.

Jira Coding Agent creates a separate session per repository. Rovo Dev in Jira can spin up and orchestrate several cloud sessions from a work item. The standalone Rovo Dev product is reaching end of life, with its CLI becoming the Atlassian Rovo CLI, according to Atlassian’s licensing page.

A work item reaches an agent through the Assignee field, an @mention, a workflow transition or a board column, using its summary and description as context. Assignee is a single-user field, but an @mention does not assign the item. The remote agent API starts another parallel task when someone assigns or mentions the same agent on the same work item again.

Cursor, GitHub Copilot and a Claude agent for Jira in beta can be assigned Jira work. Atlassian’s pages disagree about Codex availability. Agents consume Rovo credits. From December 3, 2026, the allowances are 25, 70 or 150 credits per user a month on Standard, Premium and Enterprise. Additional use costs $0.01 per credit.

Rovo MCP exposes Jira operations without a claim tool

Jira has a remote MCP connection through Atlassian’s Rovo MCP server at https://mcp.atlassian.com/v2/mcp. Atlassian recommends OAuth 2.1. API tokens also work when an admin enables that authentication method.

The legacy SSE endpoint stopped being supported after June 30, 2026.

The published Jira tools include getJiraIssue, searchJiraIssuesUsingJql, createJiraIssue, editJiraIssue, transitionJiraIssue and createJiraIssueLink. There is no claim or lock tool. Searching for an issue and then editing its Assignee does not give a client a documented exclusive-claim operation.

Neither documents exclusive claims or dependency gates

A single assignment field does not make agent execution exclusive. Linear has a single delegate, but documents no restriction preventing another agent from being mentioned or opening a session on that issue. Jira explicitly starts parallel tasks on repeated assignment or mention of the same agent.

Dependency links also need a separate decision about execution. Linear supports blocked and blocking relations, moving the relation under related when the blocker resolves. Its issue-property docs do not describe blocking as a gate on delegation or third-party agents.

Jira supports blocks and is blocked by links, but does not document those links preventing assignment or an agent trigger. A remote agent needs extra requests to fetch linked work items. Your execution system therefore needs its own dependency check if blocked work must stay out of agents’ hands.

In these documented tracker flows, a person or rule pushes each issue to an agent. Atlassian points to OpenAI’s Symphony to watch a Jira project, find eligible work and coordinate execution, with Jira remaining the system of record. That places the queueing responsibility in a separate component.

An agent-first queue enforces who can take the next issue

Retasc is a work queue and issue tracker that agents pull from over MCP. An agent calls next_issue to claim the highest effective-priority issue that is unblocked and unclaimed. A claim is one server transaction, so two agents can never hold the same issue. The remote endpoint is https://mcp.retasc.com/mcp over streamable HTTP. Each agent acts under a key a signed-in person minted, so its actions are attributed to that person and its runtime. The person comes from GitHub or Google sign-in, never from an identity the agent supplies.

Blocked work is never handed out. When a blocker closes, its dependents join the queue at once. A blocker inherits the urgency of the work it holds up, so a chore blocking an urgent feature jumps the queue.

Each claim is a 30-minute lease renewed by heartbeat or checkpoint. When the process dies, the laptop closes or the terminal is killed, the reclaimer returns the issue to the queue once renewal stops. Every claim hands over the last checkpoint, which stays on the issue and records what is done, what is next and the gotchas.

Every issue an agent creates must declare its blockers or say in one line why it depends on nothing. When that declaration is missing, the refusal suggests similar or mentioned open issues to link. A create that looks like open work is refused until the agent names the look-alike issues it read. The check includes issue ids mentioned in the new title or body. It also compares the title with open issue titles that share rare words. The agent can comment on or claim the existing issue, or re-file naming the issues it read. Review is dispatched to a different person’s agent. The server refuses to let whoever sent the work to review claim or accept that review. An agent cannot move work to review without naming the branch, PR or commit, unless it has already left a checkpoint. The reviewer receives that handoff note.

Imports move a project without creating a live tracker sync

You can import a Linear, Jira, ClickUp, Asana or Shortcut project into a new Retasc project. The Linear and Jira importers read the source and never write back.

Retasc reads Linear only when you run the import. It does not keep the API key. Retasc does not sync after an import. Re-importing updates the existing issues by source id. The same absence of ongoing sync applies to Jira.

GitHub and GitLab use a different intake path through verified webhooks. A member must approve the exact issue text before an agent can receive it, and a source edit voids that approval. When the connector has an access token, marking the issue done or canceled in Retasc closes the source issue.

An org owner can export Retasc at any time. The JSON export includes projects, issues, comments, labels, members and attachment files. The CSV export lists the issues.

The choice depends on who dispatches the work

Choose Linear or Jira when a person should delegate each issue inside the tracker where the team already plans. Linear gives you a separate agent delegate and a session lifecycle. Jira gives you assignment, mentions and workflow triggers, plus its own coding-agent options.

Choose a queue when agents should take the next eligible issue themselves across Claude Code, Codex, Cursor, OpenCode, Gemini CLI or other MCP clients. After a lease lapses, Retasc keeps the issue in its owner’s lane so that person’s next agent can resume from the checkpoint on any runtime. If nobody returns for 7 days, the issue returns to the shared pool.

Retasc bills per action. There are no seats, and members are free and unlimited. Dispatch costs $0.007 per action. Every other action costs less. Creating an issue, editing its body, recording a checkpoint or commenting costs $0.005 per action. A status, priority or title change costs $0.003. A heartbeat costs $0.0001. A new org gets $10 of usage at signup, no card.

Run npx @retasc/cli@latest bind in the folder you want to connect. If you are not signed in, the command asks for browser approval once. The command gives the folder an agent key and wires the MCP server for Claude Code, Codex, Cursor, OpenCode, Gemini CLI, Grok and Pi on the machine. Start or restart the agent in that folder.

Questions that come up

Run one command in your repo so your agents pull the next unblocked issue from Retasc, whether you import from Linear or Jira or start fresh. A new org gets $10 of usage at signup, no card.