Atlassian Rovo

Rovo is bundled with the seat and metered by the run.

Atlassian gave Rovo to every Jira and Confluence customer, then put a credit meter on it that bills by default from December 3, 2026. Here’s how that shapes its agents, and what to check before a rule runs one unattended.

Image: Atlassian · Rovo product page

At a glance#

What it does: Search, chat and agents across Jira, Confluence and Jira Service Management. Agents create and update work items, review pages, run inside automation rules, and since May 2026 can be assigned Jira work like a teammate.
Price: Included in Standard, Premium and Enterprise plans. Each user adds 25, 70 or 150 credits a month to one shared pool (250, 700 or 1,500 in the Teamwork and Service Collections). A chat answer or basic agent run costs 10 credits; search is free. From December 3, 2026, extra usage bills by default at $0.01 a credit.
Where it lives: Jira, Confluence and Jira Service Management chat and editors, Automation rules, Rovo Studio (where agents are built), and an MCP server that lets outside AI tools read and write Atlassian data.
Model: Rovo Chat runs on OpenAI GPT, Anthropic Claude, Google Gemini and open-source Llama models; Rovo Agents list OpenAI and Google. You can’t choose the model unless you build an advanced agent, which costs variable credits.
Takes actions?: Yes. In chat, tools that change data ask for confirmation first, and Jira bulk changes stop at 20 work items. In an automation rule the agent can write directly with no confirmation unless the builder limits it to read actions.
Claimed results: 5 million monthly users (February 2026). 14 million Rovo-assisted actions in one month (May 2026). Over 80% of the Fortune 500 use it, assisted actions rose 50% quarter on quarter, and Rovo customers grow spend twice as fast as others (August 2026).
Launched: Announced May 2024. On sale October 9, 2024 at $20 per user per month. Folded into Premium and Enterprise plans from April 2025 and into every Standard plan by the end of 2025.

Why it matters#

Atlassian sold Rovo as a $20 add-on for six months, then gave it to every Standard, Premium and Enterprise customer. In its place came a meter: credits, pooled across the company, with a chat answer or an agent run costing 10. On December 3, 2026 that meter starts billing, by default, at a cent a credit.

Rovo comes with the seat, and every run draws from a shared credit pool that bills extra use by default from December 3, 2026. Automation rules are where runs add up. This page covers the pricing history, the agent that asks first in chat but can write directly in a rule, model choice as a paid tier, and the Teamwork Graph Atlassian opened to outside agents.

How it works#

What happens when a request reaches Atlassian Rovo.

Ask
A person asks in Rovo Chat inside Jira or Confluence, or @mentions an agent: “Create a Jira task for this research summary.” The request spends 10 credits from the company’s pool. Searching is free.
Gather context
Rovo pulls what it needs from the Teamwork Graph, Atlassian’s map of how work items, pages, people and goals connect. It sends the request to a model Atlassian chose, from OpenAI, Google or others.
Fill the gaps
The agent asks for anything a Jira task needs that the request left out, such as the project or the issue type.
Confirm
Tools that may change data ask for confirmation before they run. A bulk change in Jira stops at 20 work items at a time.
Act as someone
The task is created as the person, or as the agent’s own account if the builder chose that in Rovo Studio. Actions in connected tools like Slack appear as if the person did them.
Move it to a rule
A builder schedules the same agent in an Automation rule, say every Monday at 9:00 AM. The person drops out. Each run spends credits and automation steps, the audit log records it, and a review page shows where chat users get stuck.

The decisions worth studying#

What they chose, why, and what it costs.

1It gave Rovo away, then put a meter on it

Rovo went on sale on October 9, 2024 at $20 per user per month. On April 9, 2025, Atlassian said it would fold Rovo into Premium and Enterprise plans, with Standard to follow, and every paid site had it by the end of 2025. Chief executive Mike Cannon-Brookes gave simplicity as the reason: buying everything separately was hard to explain.

What replaced the price is a meter. Each Jira or Confluence user adds 25, 70 or 150 credits a month to one pool the whole company shares. A chat answer or a basic agent run costs 10 credits; search is free. On September 1, 2026, Atlassian announced that from December 3, 2026 extra usage bills by default at $0.01 a credit. Admins can cap it or turn it off; with it off, billable interactions pause until the month resets.

My read: free with the seat is how you get 80% of the Fortune 500 using it. A meter that bills by default is how you get paid.

2The same agent asks first in chat and writes unattended in a rule

In Rovo Chat, Atlassian’s tools guide says an agent asks for confirmation before running tools that may change data. Jira bulk changes stop at 20 work items at a time.

Put the same agent in an Automation rule and the person drops out. The automation guide says agents there take actions directly and can be set up without confirming each one. The tools guide still says an automated agent only returns text. Atlassian’s own recording shows a warranty-audit agent scheduled for every Monday at 9:00 AM, with a second step filing its answer as a Jira task. No step asks a person.

Atlassian gives the builder two real controls. A Read actions only switch on the Use agent step. And a choice of running the agent as the person who set up the rule or as its own account with narrow access. The guide warns that on a person’s account the rule can reach everything that person can, whatever its tools. Each run also spends credits while nobody watches.

Automation rule: When: Scheduled, Every week on Mon at 9:00 AM; Then: Use agent, Laptop Warranty Audit Manager; And: Create a new Task in IT Support. The Create work item panel: Space IT Support (ITS), Issue type Task, Summary “Laptop Warranty Audit Review”, Description {{agentResponse}}, Create work item as Current Rule Actor, “The name displayed as the reporter”. — The rule runs the agent every Monday and files its answer as a task created as Current Rule Actor. No confirmation step sits between them, and each run spends credits. Atlassian recording · 1:27

3Picking the model is a paid tier

Rovo doesn’t let you choose the model. Atlassian’s transparency notes say Rovo Chat runs on OpenAI, Anthropic and Google models plus open-source Llama models, and Rovo Agents on OpenAI and Google. Atlassian decides which one runs. Since April 2026, Gemini 3 Flash powers certain Rovo capabilities under an expanded Google Cloud deal, and Atlassian is building training and inference infrastructure on Google Cloud.

The one way to pick a model is to build an advanced agent with model selection. That moves the agent from the 10-credit tier to variable credits, alongside Think Deeper, agents in Jira and code review.

My read: this is the mirror image of Fin, which trained its own model to control cost. Atlassian rents every model, keeps the routing to itself, and charges more when you want a say. That keeps Atlassian’s cost per run as its own lever. The risk is yours to measure. The same agent can answer differently as the routing changes, and the docs don’t say which model answered.

4It opened its map of work to other companies’ agents

Atlassian’s real asset is the Teamwork Graph, its map of how people, work items, pages and goals connect: 150 billion connections as of May 2026. At Team ’26 Atlassian opened it. A command-line tool with 300-plus commands and an MCP server let outside agents read and write Atlassian data. MCP v2 went generally available on September 8, 2026, and on September 29 admins got a data security policy control over it.

By August 2026, monthly users of the MCP server and the command-line tool passed 1 million. Atlassian says agents grounded in the graph give up to 44% more accurate answers using 48% fewer tokens. Enriched graph API calls cost 1 to 10 credits each.

My read: Atlassian wants to be the context every agent pays to read. Rovo is one customer of the graph among many. For buyers, the MCP server is governed apart from Rovo: read, write and search permissions per app, and audit events that may name a service account instead of a person.

What to watch#

The meter turns on by default
On December 3, 2026, extra usage starts billing unless an admin caps it or turns it off. Allowances from every product pour into one pool, so one heavy scheduled agent can drain the credits of every team. Automation rules get their own meter, at $0.50 per 1,000 extra steps.
Two help pages disagree
The tools guide says an agent in a rule only returns text. The automation guide says it writes directly. Test what your site does before the first unattended run.
Who can read a failed run
The configuration guide says the agent’s creator or an admin can open Debug response. Atlassian’s debugging guide limits it to rules the reader created. Check on a real site before you rely on either.
Rovo Dev is being folded in
The standalone coding agent is reaching end of life, with dates still to be announced. Its code reviewer moves to Rovo credits on December 3, 2026.

What it means for your team#

What to decide before you ship an agent like Atlassian Rovo.

Finance
Before December 3, 2026, set the extra-usage cap in Atlassian Administration or turn it off, and decide who pays for credits a scheduled agent spends. Rovo credits and automation steps are two separate meters.
Security
Give every automated agent its own account with narrow access instead of the builder’s account. Set the Use agent step to Read actions only unless the rule must write.
Product
Treat moving an agent from chat to a rule as a new release. Re-check the acting account, the allowed writes, the destination and the cost per run before it fires.
Platform
Build scheduled runs so a retry cannot file the same task twice. Atlassian retries some failed rules from the failed step for up to seven days and documents no general undo.

Steal this

Price the agent in one unit, publish the rate, and let admins cap it before billing starts.
Review permissions again whenever an agent moves from chat to a schedule. Nobody confirms a scheduled run.
Let the builder choose whether the agent acts as the person or as its own account.

Reference#

The evidence behind this teardown: recordings, sessions, what Atlassian Rovo gets right and wrong, and every source.

Walkthrough · 5 frames

From warranty-audit instructions to an unassigned Jira task#

Frames from Atlassian’s “Rovo Agent in Automation” recording in the Atlassian Community post “Your Assets data at your fingertips with Rovo”. A vendor recording, not a live test; the post doesn’t date the recording.

Step 1 of 5: The builder writes the audit agent’s instructions
Image: Atlassian recording · 0:24
1. The builder writes what the agent does “On each run”: find laptops whose warranty expires in the next 4 weeks.
2. The agent’s output is text: a “succinct markdown summary” with a count line and a table.
Step 2 of 5: The builder schedules the agent for Mondays
Image: Atlassian recording · 0:49
1. The builder types the trigger in plain words: “Every Monday at 9am”.
2. The rule gets its own instructions, including what to return “if no laptops are found”.
3. The builder picks where the agent works: Jira, with Global scope rather than one space.
Step 3 of 5: A separate step files the agent’s answer
Image: Atlassian recording · 1:27
1. The rule runs the Laptop Warranty Audit Manager every week on Monday at 9:00 AM; no person starts the run.
2. A native Create work item step puts the agent’s answer, {{agentResponse}}, in a new task; no step asks a person first.
3. The task is created as “Current Rule Actor”; the recording doesn’t show which account the rule actually runs as.
Step 4 of 5: The audit log records the run
Image: Atlassian recording · 1:52
1. The run logged SUCCESS at 17:35:51, five seconds after a config change; the recording doesn’t show a Monday 9:00 AM run.
2. Each stage gets its own line: Scheduled, Use agent, then Create work item with “Issues created successfully”.
3. The log links the created task twice, both times as ITS-4330: one task, not two.
Step 5 of 5: The report lands as an unassigned task
Image: Atlassian recording · 1:57
1. The agent’s answer becomes the task description: “Found 4 laptops with warranty expiring in the next 4 weeks.” and a table.
2. The task is Open and Unassigned: the report exists, but nobody owns the follow-up yet.
3. Jira shows a person as reporter; the recording doesn’t show whose permissions the rule used to create the task.