Atlassian announces AMP, GA MCP Server and Rovo Work at Team '26 Europe

At its Team '26 Europe event in Amsterdam on 7 October, Atlassian announced a broad agentic-AI push across its development and collaboration tools, headlined by the Agentic Multiplayer Protocol (AMP), a generally available Atlassian MCP Server and Rovo Work, a new mode in Rovo Chat for long-running tasks. According to dev.to, whose writeup worked through the announcements on a test organisation, the keynote produced 11 distinct items — but only one carries a generally available label, while the admin switches governing external AI access to company data are already live.

AMP makes agents visible teammates

Atlassian frames AMP as its umbrella for the experiences, technologies and patterns that define how people and agents collaborate. Concretely, agents gain a presence inside Atlassian products: they appear alongside humans on canvases and whiteboards, with their own cursors, and can be registered once and then chosen to surface on a Jira board, a Confluence page or in Rovo Chat. Co-founder Mike Cannon-Brookes said the rollout begins immediately, but as dev.to notes, Atlassian has published no plan tiers, most sub-features are labelled anywhere from beta to "Soon", and there is no admin setting actually named AMP. What admins configure instead is agent identity, with agents able to run as a user or on dedicated service accounts; Atlassian's documentation suggests the agent's own account to minimise security risk.

The MCP Server is the only generally available piece

The Atlassian MCP Server — still labelled "Rovo MCP server" in the admin console — has been generally available since 8 September, per the changelog. Atlassian counts 220+ tools in its MCP announcement (the platform post says 200+) and claims up to 25% fewer tokens in internal benchmarking on Claude models.

dev.to inspected the permissions page on a test organisation and found 36 toolsets, 19 tagged New, with every read, write and search toolset already allowed — including eight new write toolsets such as write_projects, write_goals and write_teams. New read toolsets included read_loom and read_talent, which reach Loom meeting recordings and Talent headcount metrics. Delete and Manage were blocked, notable because delete_jira can permanently remove issues, comments and attachments.

Atlassian's permissions documentation states that new permissions do not require manual review by administrators. Vetting write access takes a deliberate step: switching Write from "Allow all" to "Allow some" toolsets, which also turns off "Allow new write toolsets by default" — the very default Sami Shaik flagged publicly two days before the keynote.

The dev.to report adds caveats: domain allowlists and the MCP control inside data security policies apply only to OAuth connections, so tools using API tokens or the beta enterprise managed authentication bypass them. Every tool call does land in the audit log on all tiers, and on 1 March 2027 any remaining v1 usage automatically moves to v2 tools. The nearby Teamwork Graph CLI fared worse in testing, with all 32 of its permissions allowed by default, including seven delete permissions.

Rovo Work exists on paper only, for now

Rovo Work is described as tackling complex, multi-step work, with tasks that can run for hours inside a secure sandbox that admins govern. dev.to found no status label, no documentation page and no Work option in Rovo Chat on a test site with beta features enabled. Pricing is equally open: the licensing FAQ lists Work mode among premium AI interactions with variable pricing, and extra-usage billing for Rovo credits starts on 3 December 2026 at a listed $0.01 per credit — enabled by default per Atlassian's docs, though the test organisation still required opting in.

EU inference, code context and agent inventories trail behind

Atlassian says LLM processing will be restricted to models hosted within the EU and is rolling out, but dev.to could find neither a doc page nor a setting for it; data residency governs where data is stored, not where the model runs. Code context is in open beta and auto-enables for paid Bitbucket workspaces connected to a Jira site with Rovo — but indexed code is stored and processed in the United States, without data residency support. A unified inventory of agent accounts and non-human identities is marked "Soon".

Why it matters

Atlassian is positioning its suite as a shared workspace where AI agents operate shoulder to shoulder with employees, and the plumbing for that — MCP toolsets, agent identities, hour-long task execution — is landing before the headline features are finished. The immediate consequence is operational: in at least one test organisation, broad read and write access to Jira, Confluence, Loom and Talent data was switched on by default, and assumed safeguards such as allowlists and data residency do not cover every connection type. Teams running Rovo should audit their MCP and Teamwork Graph permissions now, and budget for credit-based billing before December.