Rovo Dev is Atlassian’s coding agent, and the headline number is real but needs an asterisk: Atlassian’s CLI sits at #1 on the SWE-bench Full leaderboard with a 41.98% resolve rate across all 2,294 tasks. That’s a legitimate top spot — but it’s SWE-bench Full, not SWE-bench Verified, which is the harder-to-game subset most other vendors (Claude Code, Devin) quote. So Rovo Dev is genuinely good, and the marketing is also doing a little sleight of hand. This page is the operator’s read on the product itself: what it is, how to install it, what it can and can’t do, and how it actually compares to the other agents you’d evaluate — with pricing and the connector story kept as links, not re-argued here.

What is Rovo Dev, exactly?

Rovo Dev is a separate product from Rovo. That distinction is the single most important thing to get right, so start here: “Rovo” is the AI layer — agents, chat, search — living inside Jira and Confluence, metered by a credit allowance included with your paid plan. Rovo Dev is a standalone, separately-priced agentic coding tool that reads and writes actual code. If you’ve been reading our Rovo DevOps integration page, this is “Layer 2” from that map, covered here at full depth.

It ships in three surfaces, and they’re the same agent behind different front-ends:

  • The CLI — an extension of the Atlassian CLI (ACLI). You describe a task in your terminal and the agent reads the codebase, proposes multi-file changes, and runs commands.
  • The IDE extension — Rovo Dev went generally available in VS Code in early 2026, and also runs in VS Code-compatible IDEs like Cursor. Same agent, in your editor, with Jira/Confluence/Bitbucket/GitHub context pulled in through Atlassian’s Teamwork Graph.
  • AI code review — Rovo Dev reviews pull requests and, crucially, checks them against the acceptance criteria in the linked Jira issue.

The through-line — the thing that isn’t just “another coding agent” — is that all three can act on your Jira work items and create Confluence content directly, no copy-paste. That’s the Atlassian-native part, and it’s the whole reason to pick this over a general-purpose agent.

How do you install and run the Rovo Dev CLI?

Three commands, once the prerequisites are in place. Per Atlassian’s install docs, it runs on macOS, Linux, and Windows, and you need a site with Rovo Dev activated plus a Rovo Dev credit allocation before anything works.

  1. Install (or update) ACLI. Rovo Dev CLI is an extension of the Atlassian CLI, so you install ACLI first using the OS-specific instructions, or update it if you already have it.
  2. Create an API token. In your Atlassian profile, generate an API token — the installer offers an automatic link, or you create one manually with the required scopes (a long list spanning Jira, Confluence, and Bitbucket read/write). Copy it to your clipboard.
  3. Authorize and run:
acli rovodev auth login
acli rovodev run

acli rovodev auth login authorizes the CLI with your Atlassian account; acli rovodev run opens an interactive session. That’s it — you’re talking to the agent.

One gotcha worth flagging before you evaluate: the CLI’s fuller feature set requires a paid Rovo Dev Standard subscription, so budget for the $20/developer line (there’s a 30-day free trial to kick the tires first).

What can Rovo Dev actually do — and what can’t it?

Per the Use Rovo Dev CLI docs, the agent covers the now-standard coding-agent surface plus a few genuinely differentiated pieces:

  • Code work — generate, review, and refactor code; analyze a codebase; write documentation; debug issues; create unit tests.
  • Atlassian actions — action Jira work items and create Confluence content without leaving the terminal. This is the differentiator.
  • Sessions and memory — sessions keep history and context across interactions, with user memory and project memory to retain preferences and project facts.
  • Extensibility — Agent Skills for customizing task behavior, MCP server connections to pull in external data and tools, and subagents with their own system prompts and tool configs.
  • Modes — interactive, server mode for continuous operation, and a worktree mode for isolated Git operations with cleanup.
  • Model switching — you can switch between multiple LLMs, with credit multipliers per model (i.e., the fancier model burns credits faster).

What it can’t do, or where the edges are: it’s built around the Atlassian ecosystem, so its unique value collapses if your work doesn’t live in Jira/Confluence/Bitbucket — outside that, it’s a competent-but-not-special agent. And the meter is always running: every capable model call draws down credits at a multiplier, which is a real constraint on how loosely you can let it churn. For what a credit is and how the two separate meters work, see Rovo pricing and credits explained.

Is Rovo Dev really #1 on SWE-bench?

Yes, with the asterisk from the top of this page. Atlassian’s published claim is a 41.98% resolve rate across 2,294 tasks, which is #1 on the SWE-bench Full leaderboard. SWE-bench is the standard benchmark for agents resolving real GitHub issues, so this isn’t a vanity metric — it’s a hard, real-world test.

The operator’s caveat: that’s SWE-bench Full (all 2,294 tasks), whereas the numbers you see quoted for Claude Code, Devin, and most others are usually SWE-bench Verified — a human-curated subset of tasks confirmed to be solvable and well-specified. The two aren’t directly comparable, and “we’re #1 on Full” isn’t the same claim as “we beat Claude Code on Verified.” So read the benchmark as Rovo Dev is a strong, top-tier agent — which it is — not as proof it’s the best coding agent you can buy. Benchmark leadership on the metric a vendor chose to report is always worth one raised eyebrow.

Rovo Dev vs Cursor, Claude Code, and GitHub Copilot

Put Rovo Dev in the right bracket first. It is not an in-editor autocomplete like GitHub Copilot’s core product — it’s an autonomous agent, so its real peers are Claude Code (terminal agent) and Cursor’s agent mode (IDE agent). Feature-for-feature at the agent layer, they rhyme: read the repo, multi-file edits, run commands, MCP support, subagents.

The differences that matter for a buying decision:

  • vs Claude Code / Cursor — the delta is native Atlassian context. Rovo Dev can pull a Jira ticket’s acceptance criteria, act on the work item, and open a Bitbucket PR as part of the same loop, where Cursor or Claude Code would need an MCP bridge to get the same context. If your team lives in Jira, that’s a real, daily time-saver. If it doesn’t, you’re choosing on raw agent quality and price, and the Atlassian tie-in stops being a reason.
  • vs GitHub Copilot — different jobs, not a head-to-head. Copilot writes code inline as you type; Rovo Dev is the agent that goes and does a whole task. Atlassian even ships a “Rovo for GitHub Copilot” extension so the two run together. The full breakdown of Rovo vs the Microsoft Copilots (and the three-different-Copilots confusion) is in Rovo vs Copilot for Jira teams.

The honest summary: Rovo Dev’s moat is the Atlassian graph, not the model. Pick it when your workflow is Jira-native and you want the agent to close the loop from ticket to PR; look harder at Claude Code or Cursor when the codebase, not the ticket, is the center of gravity.

What does Rovo Dev cost?

Kept short because it’s supporting, not the point: $20 per developer per month for Rovo Dev Standard, with a 30-day free trial, and an allocation of 2,000 Rovo Dev credits per developer per month; overage is billed at $0.01 per credit (Rovo Dev pricing, how billing works). The one thing to internalize: these credits are entirely separate from the standard Rovo allowance included with your Jira plan — different pool, different meter, its own bill. To size a team’s monthly draw before committing, run it through our Rovo credit cost estimator, and for the full two-meter model see Rovo pricing and credits explained.

Should you use it?

If your engineering team already runs on Jira and Bitbucket, Rovo Dev is worth the 30-day trial on the strength of the native context alone — the ticket-to-PR loop is the part generic agents can’t replicate without plumbing. If you’re a heavy Jira shop evaluating agents, start it on one repo’s pull requests (the AI code review, checked against linked-issue acceptance criteria, is the lowest-risk on-ramp) before rolling the CLI out broadly. And if you’re not deep in the Atlassian ecosystem, the honest answer is that Rovo Dev’s headline strength doesn’t apply to you, and you should evaluate it as one strong agent among several rather than the obvious pick its #1-on-SWE-bench banner implies.

We build Rovo agents at Crosstown Tech — AI Playbooks for Rovo is ours — which is why we keep working notes like this on how the platform’s pieces actually fit together.

Want one built rather than built by you? We take Rovo agents, Forge apps and Jira customizations from spec to deployed. Configure what you need and see the price and timeline instantly →