Atlassian Rovo Agents, MCP Server, and AMP: The Complete Guide for Agentic AI Architects (2026)

A Jira ticket sits in “In Progress” for three days. The developer assigned to it has been in back-to-back meetings, the Confluence design spec was updated twice since they last read it, and the Bitbucket branch they started from is now four commits behind main. Nobody notices until standup.

Now imagine the same ticket, but an AI agent is watching. It reads the full comment thread, checks the latest Confluence spec, notices the branch drift, flags the blocker in a comment, and — if your admin has enabled it — drafts a fix and opens a pull request. The developer reviews the PR after their meetings, not the raw problem.

That is the shift Atlassian announced at Team ’26 Europe in October 2026: agents are no longer assistants you visit in a chat window. They are teammates with identity, presence, and permissions that show up alongside humans in Jira boards, Confluence pages, and Bitbucket repositories. The technical machinery behind this — a rebuilt MCP server with 220+ tools processing over 15 million calls per day, a new Agentic Multiplayer Protocol (AMP) for human-agent collaboration, and an expanded OpenAI partnership layering GPT-6 Astra into a multi-model routing architecture — represents one of the most complete enterprise agent platforms shipping today.

For agentic AI architects and RPA professionals, this matters because Atlassian is not building another chatbot. It is turning the tool where 300,000+ organizations already track their work into a first-class agent orchestration surface — and it is doing it with an open MCP architecture that connects to Claude, ChatGPT, Gemini, GitHub Copilot, Cursor, and any MCP-compatible client.

This guide breaks down every layer of the new stack: what shipped, what is in beta, what it costs, and how it fits into the broader agentic AI landscape that RPA teams are navigating right now.

What Atlassian Announced at Team ’26 Europe: The Full Picture

The Team ’26 Europe conference (October 6–7, 2026) landed five interconnected announcements that, taken together, redefine how agents operate inside enterprise work management:

Each of these layers reinforces the others. The MCP server provides the tool surface. AMP provides the collaboration model. Rovo Work provides the execution runtime. The OpenAI partnership (alongside existing Anthropic Claude and Google Gemini integrations) provides the reasoning layer. And agent accounts provide the identity and governance layer that enterprise buyers demand.

Let us examine each one.

The Rebuilt Atlassian MCP Server: 220+ Tools and 15 Million Daily Calls

The MCP server is the backbone of Atlassian’s agentic strategy. Rebuilt from the ground up since the original December 2025 ChatGPT integration, it is now the single access point through which any AI agent — first-party or third-party — reads, writes, and acts on Atlassian data.

Architecture and Endpoint

The server is cloud-hosted by Atlassian at https://mcp.atlassian.com/v2/mcp (the legacy /v1/mcp and /sse endpoints remain supported, with v1 tools migrating to v2 automatically on March 1, 2027). Authentication uses OAuth 2.1 with PKCE by default, with API token authentication available for headless and CI/CD use cases — though an org admin must explicitly enable API token auth first.

A key architectural decision is progressive tool discovery. Rather than dumping all 220+ tool definitions into a client’s context window on connection, the server advertises a small set of primary tools and exposes the rest through a dedicated discover tool. This conserves context window budget (critical for cost-conscious agentic deployments) and allows Atlassian to add new tools without requiring clients to reconnect.

Anthropic’s head of enterprise product, Scott White, called it “one of the most-used enterprise MCP integrations on Claude” — and the 25% token reduction claim was benchmarked specifically on Claude models.

Tool Coverage by Product

The rebuilt server spans Atlassian’s full product suite:

The Teamwork Graph deserves special attention. Described by Atlassian as a layer mapping more than 150 billion objects and relationships across Jira, Confluence, Bitbucket, Loom, and 75+ third-party tools, it is what gives agents pre-linked context on the first turn. When an agent picks up a Jira ticket, the Teamwork Graph can surface the related Confluence spec, the Bitbucket branch, the Loom walkthrough, and the team members who last touched each artifact — without the agent having to make separate discovery calls.

Security and Admin Controls

Enterprise governance is where Atlassian’s MCP server differentiates from most open-source MCP implementations:

• Permission inheritance: Every tool call runs under the authenticated user’s existing Atlassian permissions. Agents never get more access than the user already holds.
• Risk-tiered actions: Read, write, search, and destructive actions are separated into distinct permission groups. Clients can require human confirmation for higher-risk steps.
• Atlassian Guard integration: Data Security Policies and Data Loss Prevention controls apply to MCP connections on Guard Standard and above.
• Admin Console controls: Domain allowlists, IP allowlists, per-scope read/write/search permission toggles, and organization-wide audit logging.
• Full audit trail: Every agent action is logged — on all plan tiers, not just Enterprise.

One nuance worth noting for security teams: the domain allowlist and MCP data-security controls apply only to OAuth connections. API-token and enterprise managed authentication (currently in beta) bypass them. This is a known gap that architects should factor into their security posture.

Supported Clients

The MCP server works with any MCP-compatible client, including: OpenAI ChatGPT, Claude (claude.ai, Desktop, and Code), Cursor, VS Code (GitHub Copilot), GitHub Copilot CLI, Google Gemini, Amazon Quick Suite, Kiro, Docker, and Windsurf. One-click install is available for Cursor, VS Code, ChatGPT, and Claude. Local clients use the mcp-remote proxy (Node.js v18+ required).

The open-source server repository (atlassian/atlassian-mcp-server) carries an Apache-2.0 license and has accumulated over 1,000 GitHub stars.

AMP: The Agentic Multiplayer Protocol

If the MCP server is the how of agent access, AMP is the where of agent presence. Atlassian describes AMP as “the collection of human-centric experiences, technologies, and patterns that shape how humans and agents work together.”

In practical terms, AMP means:

• Agents appear in real-time presence bars and cursors alongside human teammates — on Jira boards, Confluence pages, and Rovo Chat.
• Registration once, deployment anywhere: You register an agent in Rovo Studio and then choose where it appears — a specific Jira board, a Confluence space, or Rovo Chat.
• Identity options: Agents can run as a user (inheriting that user’s permissions) or use dedicated service accounts with scoped permissions.

This is a conceptual shift from how most enterprise agent platforms work today. In platforms like Salesforce Agentforce, agents are invoked through dedicated agent surfaces. In Atlassian’s vision, agents are co-present — they show up in the same UI where humans are already working, with the same visual indicators of activity.

The implications for RPA teams are significant. Traditional RPA bots operate behind the scenes, processing queues and triggering workflows in the background. AMP positions agents as visible, accountable teammates whose actions are transparent to everyone on the board. This is closer to the “attended automation” model that RPA architects have long sought — but with the reasoning capabilities of frontier LLMs instead of rule-based scripts.

Rovo Work: Multi-Step Agent Tasks That Run for Hours

Rovo Work is a new Rovo Chat mode designed for long-running, multi-step tasks that can execute for hours in an admin-governed sandbox. While details remain limited (it was not visible on test sites at the time of Team ’26 Europe), Atlassian positions it as the bridge between simple chat-based agent interactions and full autonomous workflow execution.

Key characteristics:

• Tasks run in a sandboxed environment with admin-defined boundaries
• Falls under premium AI interactions with variable pricing based on compute consumed
• Designed for complex, multi-tool workflows — triaging a backlog, auditing a Confluence space for outdated content, generating release notes from a sprint’s worth of commits

For agentic AI architects building on Atlassian, Rovo Work represents the execution layer that sits between ad-hoc agent queries (which consume 10 Rovo credits per request) and fully autonomous agent workflows (still on the roadmap). It is where the comparison to UiPath’s long-running orchestrated processes or Automation Anywhere’s agentic process automation becomes most direct.

The OpenAI Partnership: GPT-6 Astra Inside a Multi-Model Router

The expanded OpenAI partnership adds GPT-6 Astra (released September 3, 2026) and the GPT-5.6 series to Rovo’s model palette. But the architecturally significant detail is that Atlassian is not making this an exclusive arrangement.

Rovo’s Multi-Model Routing

Rovo uses an internal AI gateway that dynamically routes tasks across providers based on three factors: capability (which model handles this task class best), speed (latency requirements), and cost (token economics). GPT-6 Astra is available but not the default model — Atlassian Intelligence already runs on Anthropic Claude, Google Gemini 3 Flash, and other providers.

The VentureBeat reporting on this partnership noted that Anthropic’s Claude Fable 5.1 scored higher than Astra on the Artificial Analysis Intelligence Index — a measure OpenAI itself publishes. Atlassian’s multi-model approach means it can route each request to whichever model offers the best quality-cost tradeoff for that specific task.

GPT-6 Astra: Capabilities and Cautions

GPT-6 Astra is priced at $10 per million input tokens and $50 per million output tokens. It is OpenAI’s first model to reach the “Critical” cybersecurity threshold under its Preparedness Framework, scoring 100% on one exploit benchmark and surfacing two previously unknown zero-day vulnerabilities during testing.

For enterprise deployments, important cautions apply:

• Safety checks can “slow, pause, or stop legitimate work” — including defensive security tasks
• OpenAI acknowledges that Astra’s written reasoning is “harder to monitor than its predecessor’s,” which it considers “a serious regression”
• Astra currently refuses advanced offensive tasks (e.g., writing proof-of-concept exploits); OpenAI plans to loosen this for vetted defensive workflows
• Enterprise access is off by default until an admin explicitly enables it
• Zero Data Retention is available for “eligible” API customers, though what “eligible” means is not fully defined

Agent Accounts and Identity: The Governance Layer

Perhaps the most architecturally significant announcement for enterprise RPA teams is the introduction of agent accounts — dedicated, persistent identities for AI agents that work like user accounts but with scoped permissions and full audit trails.

How Agent Identity Works

In Rovo Studio, administrators can configure agents with two identity models:

• User’s account: The agent inherits the permissions of the human who invoked it. Actions are attributed to that user.
• Agent’s account: The agent operates under a dedicated service account with its own scoped permissions. Actions are attributed to the agent identity.

Agents can be assigned Jira work items, @mentioned in comments, and move through Jira workflow states just like human teammates. They show up as assignees in sprint boards and their session status (stuck, waiting, or complete) is visible to the team.

This is a direct answer to one of the thorniest governance questions in enterprise agentic AI: who did what? When an agent modifies a Confluence page, transitions a Jira ticket, or opens a Bitbucket PR, the audit trail shows which agent acted, under which identity model, and with which permissions — at every plan tier.

Cross-Vendor Identity: The Unsolved Problem

Analyst coverage from Futurum Group flags an important limitation: when a workflow spans Rovo and an outside orchestrator (such as Gemini Enterprise or Microsoft’s Agent Framework), the identity model breaks down. Which system’s identity governs the cross-boundary action? This remains unresolved — and it is precisely the kind of interoperability challenge that the Agentic AI Foundation (AAIF) and the A2A protocol are designed to address.

Agents in Action: Jira and Confluence Use Cases

The practical impact shows most clearly in specific use cases that are now generally available or in open beta:

Jira: Agents as Assignees

• Coding agent assignment (GA): Engineering teams can assign Jira work items to Claude Code, Cursor, and GitHub Copilot. The agent reads the ticket context, checks related artifacts via the Teamwork Graph, and produces a PR for human review.
• Built-in Jira Coding Agent (GA): Included in every paid Jira plan at no extra cost. Turns work items into pull requests ready for review.
• Automation-triggered agent routing: An automation rule builder lets admins route specific work types (bug fixes, remediation tickets) to agents automatically when they enter a designated status.
• DX AI cost management: A reporting layer maps token consumption and agent spend data to engineering output metrics — answering the question “what did our AI agents actually produce?”

Confluence: Agents as Collaborators

• @mention agents on pages (open beta): You can tag agents from Lovable, Replit, Databricks, Gamma, and others on a Confluence page. The agent reads the page context and acts across connected tools.
• Remix with Rovo (beta): Converts selected Confluence text into charts, timelines, and infographics.
• Confluence Slides (beta): Rovo generates and edits presentations grounded in team work, presentable within Confluence.
• Create with Rovo in Jira (beta): Turns Confluence docs, meeting summaries, and email threads into Jira work items.

Cross-Product: The Teamwork Graph in Action

The most compelling use cases span multiple products:

• Meeting-to-work pipeline: A Loom recording is transcribed, decisions are extracted, and the agent drafts Jira work items and Confluence spec pages — all without a human doing the manual conversion.
• Ticket catch-up: An agent reads the full comment thread and field-level changelog of a Jira ticket, synthesizes the history, and presents a summary with key decisions highlighted.
• Safe page editing: When an agent edits a Confluence page, it diffs document versions and changes only the requested section. If someone else edits the page mid-task, the agent pauses to avoid overwrites and previews changes before publishing.

Pricing: What Rovo Agents Actually Cost

One of the strongest aspects of Atlassian’s agent strategy is the pricing model: Rovo is bundled into all paid Jira, Confluence, JSM, and Teamwork Collection cloud subscriptions at no separate per-seat fee.

Credit Allowances

Credit Consumption

Credits are pooled across the organization, reset monthly with no rollover. As of October 2026, Atlassian is not billing for overage — requests above the pool still execute. Atlassian has committed to providing at least 90 days’ notice before overage becomes billable, with an explicit opt-in required. Extra-usage billing is tentatively scheduled to start December 3, 2026, at a list price of $0.01 per credit.

For comparison, Salesforce Agentforce charges $2 per conversation for its standard agent interactions. Atlassian’s credit-based model is substantially more granular — and currently free above the included allowance.

JSM-Specific Pricing

• Rovo Customer Service: $1 per resolution (charged only when the AI fully resolves a ticket without human escalation)
• Virtual Service Agent: 1,000 assisted conversations/month included (12,000/year on annual billing), with $0.30 per conversation overage

The Ecosystem: Third-Party Agents and Extensibility

Atlassian is positioning its platform as an open agent marketplace, not a walled garden:

• Third-party coding agents on the Atlassian Marketplace: Devin, Factory, and Warp are already available. OpenAI Codex is “coming soon.”
• Confluence agent integrations (open beta): Lovable, Replit, Databricks, and Gamma agents can be @mentioned on pages.
• Forge extensibility: Developers can build custom MCP tools with the rovo:mcp module (Preview since August 14, 2026) and add MCP capabilities to Marketplace apps.
• Bidirectional MCP with Google: Rovo is reachable from within Gemini Enterprise, and Google Gemini 3 Flash powers select Rovo capabilities.
• Over 3,000 internal users: Atlassian itself uses Codex through ChatGPT Enterprise, with its DX platform measuring delivery impact.

What Is on the Roadmap (Not Yet Shipped)

Several capabilities were announced as coming but are not yet available:

• Autonomous agents that pick up Jira work items without human assignment, run tests, and support multi-agent orchestration with human checkpoints
• Assigning tasks to AI agents with progress tracking and decision recording in Jira
• Deeper Codex integration in planning and code review workflows
• Syncing local agent session history back to team boards
• Agent Sessions (closed EAP): Linking local and cloud agent runs to work items
• Interactive PR Reviews (closed EAP)
• Service Collection AI agents in Slack, Teams, and Claude (closed EAP)
• Local EU inference: Restricting LLM processing to EU-hosted models (rolling out)
• Non-human identity inventory: A centralized view of all agent identities across the org

What This Means for RPA Teams

For the audience that reads RPABOTS.WORLD — RPA developers, automation architects, and teams evaluating the move from traditional automation to agentic AI — Atlassian’s announcements carry several implications:

1. The Work Management Layer Becomes the Agent Orchestration Layer

Traditional RPA orchestrators (UiPath Orchestrator, Automation Anywhere Control Room) manage bot queues, schedules, and credentials. Atlassian is building a parallel orchestration layer — but the “queue” is your existing Jira backlog, the “schedule” is your sprint cadence, and the “credential” is an OAuth-scoped agent account. For organizations already on Atlassian, this is an agent runtime they do not need to buy separately.

2. MCP Replaces Point-to-Point Integrations

RPA teams are accustomed to building and maintaining connectors between their automation platform and enterprise tools. Atlassian’s MCP server provides a single, standardized integration point with 220+ pre-built tools, governed by existing user permissions. Instead of building a UiPath activity to read a Confluence page or an Automation Anywhere action to transition a Jira ticket, an agentic workflow can call the MCP server directly — from any MCP-compatible client.

3. Agent Visibility Changes the Trust Equation

One of the persistent challenges in RPA adoption is the “black box” problem: business users do not see what bots are doing until something goes wrong. AMP’s agent presence model — where agents appear in real-time alongside humans on Jira boards — addresses this directly. When a team can see that an agent is working on a ticket, has been stuck for 20 minutes, or has completed its task, the trust dynamic shifts from “I hope the bot ran” to “I can see what the agent is doing.”

4. Credit-Based Pricing Aligns with RPA Consumption Models

RPA pricing has moved from per-bot licensing to consumption-based models (UiPath’s per-transaction pricing, Automation Anywhere’s FTE-based model). Atlassian’s credit-based system follows the same trajectory — and with the current zero-overage policy, organizations can experiment extensively before committing to usage-based spend.

Architect’s Checklist: Evaluating Atlassian’s Agent Platform

Before adopting Rovo agents in production, agentic AI architects should evaluate these dimensions:

How Atlassian’s Agent Platform Compares

Context matters. Here is how Atlassian’s approach compares to the other major enterprise agent platforms:

Atlassian’s differentiator is the combination of open MCP architecture (any client, any model), co-present agents (not a separate agent surface), and bundled pricing (no separate agent SKU). Its weakness is the lack of native RPA or workflow automation capabilities — it depends on partners and the Forge platform for anything beyond Jira/Confluence/Bitbucket actions.

Getting Started: Connecting Your First Agent

For teams ready to experiment, the quickest path is:

1. Verify prerequisites: You need an Atlassian Cloud site on a paid plan (Standard, Premium, or Enterprise). An admin must verify the business domain — generic email domains (gmail.com, etc.) cannot enable Rovo.
2. Enable Atlassian Intelligence: An org admin enables AI features in the Admin Console.
3. Connect an MCP client: Use the one-click install for Claude Desktop, Cursor, VS Code, or ChatGPT. For other clients, configure with the endpoint https://mcp.atlassian.com/v2/mcp and complete the OAuth 2.1 flow.
4. Start with read-only: Begin with search and read operations (ticket lookups, Confluence searches, code exploration) before enabling write operations.
5. Monitor credits: Watch credit consumption in the Admin Console. Notifications trigger at 80% and 100% of your monthly allowance.

Frequently Asked Questions

Is Rovo a separate product I need to buy?

No. Rovo is bundled into all paid Jira, Confluence, JSM, and Teamwork Collection cloud subscriptions. There is no separate per-seat fee. You get a monthly credit allowance based on your plan tier (25–1,500 credits per user per month depending on plan and bundle).

Can I use Rovo agents with Claude, not just ChatGPT?

Yes. The Atlassian MCP server works with any MCP-compatible client — Claude Desktop, Claude Code, ChatGPT, Gemini, Cursor, VS Code, GitHub Copilot, and others. Atlassian’s Rovo router itself uses Claude models alongside OpenAI and Google models internally.

Does the MCP server replace the need for Jira/Confluence REST APIs in my automations?

For agentic workflows, largely yes — the MCP server provides a higher-level interface with 220+ pre-built tools that include permission inheritance and audit logging. For traditional API integrations (webhooks, scheduled syncs, custom app development), the REST APIs remain the right choice.

What happens when an agent and a human edit the same Confluence page simultaneously?

Atlassian has built collision handling into the agent workflow: when an agent detects that someone else has edited the page mid-task, it pauses to avoid overwrites and previews its changes before publishing. This is a significant improvement over naive API-based automation that can silently overwrite concurrent edits.

How does Atlassian’s agent platform compare to UiPath or Automation Anywhere for process automation?

They serve different layers. UiPath and Automation Anywhere are purpose-built automation platforms with desktop automation (RPA), workflow orchestration, and process mining. Atlassian’s agent platform is a work management layer with agentic capabilities — it excels at software development, IT service management, and knowledge work automation within Atlassian’s product suite. The two approaches complement each other: an RPA bot can trigger Jira ticket creation via the MCP server, and a Rovo agent can invoke external automation workflows. They are not direct substitutes.

Key Takeaways

• Atlassian is building the most complete “agents as teammates” model in enterprise software — with persistent identity, real-time presence, and full audit trails inside the tool where teams already work.
• The rebuilt MCP server (220+ tools, 15M+ daily calls, 25% fewer tokens) is a production-grade integration point that any MCP-compatible AI client can use, making it the widest open agent access surface among enterprise work management platforms.
• Multi-model routing (OpenAI GPT-6 Astra + Anthropic Claude + Google Gemini) means Rovo can optimize for quality, speed, and cost per task — no single-vendor lock-in on the reasoning layer.
• Bundled pricing with no current overage billing removes the experimentation barrier. Organizations can deploy agents across their entire Atlassian footprint without per-agent licensing.
• For RPA teams: Atlassian is not replacing your automation platform — it is providing a governed agent layer on top of your work management system. The MCP server is the bridge between agentic AI workflows and your Jira/Confluence data.
• Watch the gaps: Cross-vendor agent identity is unresolved, observability is dashboard-level (not production telemetry), and Rovo Work pricing and GA timing are unclear. Evaluate these before committing to production workloads.

References

1. Atlassian Blog — “Atlassian MCP brings more of your work within reach of any AI agent” (October 2026)
2. VentureBeat — “Atlassian deepens its OpenAI partnership with a spend commitment” (October 2026)
3. DEV Community — “Atlassian Team ’26 Europe announcements: AMP, MCP” (October 2026)
4. Futurum Group — “Atlassian Fuses the Agent Work Surface, Workflow, and Control Plane Into Jira” (2026)
5. CCLeaks — “Atlassian OpenAI Deal: How ChatGPT and Codex Reach Jira via MCP” (October 2026)
6. Atlassian Blog — “Teamwork Collection: Team ’26” (2026)
7. Atlassian Developer — “Atlassian Rovo MCP Overview”
8. GitHub — atlassian/atlassian-mcp-server (Apache-2.0)
9. eesel AI — “Atlassian Intelligence and Rovo pricing explained” (2026)
10. Solutions Review — “AI News for the Week of October 9” (October 2026)

Social Media Post (X/Twitter):

Atlassian just turned Jira into an agent orchestration platform. 220+ MCP tools, real-time agent presence alongside humans, multi-model routing (GPT-6 Astra + Claude + Gemini), and bundled pricing. Here’s what RPA architects need to know: [link] #AgenticAI #RPA #Atlassian #MCP

LinkedIn Post:

Atlassian’s Team ’26 Europe announcements are bigger than most coverage suggests.

They didn’t just add AI to Jira — they built a complete agent platform:

→ A rebuilt MCP server with 220+ tools handling 15M+ calls/day
→ AMP (Agentic Multiplayer Protocol) — agents appear alongside humans in real-time
→ Multi-model routing across OpenAI, Anthropic, and Google
→ Agent accounts with persistent identity and full audit trails
→ Rovo Work for multi-step tasks running hours in sandboxed environments
→ All bundled into existing paid plans — no separate agent fee

For RPA teams: your Jira backlog just became an agent queue. Your Confluence pages just became agent workspaces. And the MCP server provides a single, standardized integration point that replaces dozens of custom connectors.

The full architecture breakdown: [link]

#AgenticAI #RPA #Automation #Atlassian #MCP #EnterpriseAI