Rovo Dev

You will like it if the ticket is the unit of work: the agent sits inside Jira and Bitbucket, plans the work item, writes the code and reviews the pull request without leaving the tools your team already lives in, and the VS Code integration rides the AtlasCode extension you may already have. You will not like it if you want a fast shell; the daily trait that decides it is context from the ticket, not speed at the prompt.

Pick it for the Atlassian shop where the work is already in Jira. Pick Claude Code for the terminal-first engineer, and Devin for ticket-to-PR without the Atlassian dependency.

The risk is the meter shape. Standard includes 2,000 Rovo Dev credits a month and overage is $0.01 per credit, so a retrying agent pays per attempt, and there is no sandbox, so those attempts run in your shell with your permissions. A loop that burns credits and a loop that runs commands are the same loop here.

Set the overage cap before the pilot, not after. What it does right: an MCP server can have enable_instructions set to false, so an untrusted server cannot write into the system prompt, a prompt-injection defense most MCP clients do not offer.

Rovo Dev publishes no benchmark. The design is a thin agent over three hosted model vendors, OpenAI, Anthropic and Google, with the routing undocumented. 1. The terminal agent is an extension of the Atlassian CLI, acli, not a standalone binary, so its surface is inherited from a tool built for administration. 2. Context comes from the working tree plus Jira and Bitbucket objects, the one architectural idea here: the ticket is part of the prompt.

Edit application and verification are not described beyond the shell. The observation: the integration is the product, not the loop.

The company is Atlassian, with Jira and Bitbucket as distribution nobody else can buy. The tell is the free monthly credits rolling out to Jira Standard, Premium and Enterprise Cloud plans: bundle first, price later, the move that made every add-on in that suite a line item eventually. Pricing power comes from the seat already on the invoice, not from the agent.

No acquirer question; the likely pivot is folding it into Jira pricing so the product cannot be unbought. Position: durable, buy through the existing contract, and negotiate the credits into the renewal.

The demo knows the Jira ticket. Procurement: $20 per developer per month, $1,200 for sixty, with a 30-day free trial for the pilot. The hosted models run under zero-data-retention terms, which answers security's first question, and identity rides the Atlassian account we already administer, so SSO is already done. Headless mode exists for CI.

Onboarding is an afternoon for anyone who has used the Atlassian CLI, a day for anyone who has not. Approved with conditions: a credit budget per seat before rollout, and the overage line reviewed monthly.

Closed. No bring-your-own-key, no Ollama, model picked for me. What I can touch: ~/.rovodev/mcp.json, which acli rovodev mcp opens in my editor, with stdio, http and sse transports, and a disabledMcpServers list in ~/.rovodev/config.yml to switch servers off without deleting them. That is the whole extensibility budget, and it is a reasonable one for MCP.

A fork is impossible and a local run is impossible, so the tool exists at the vendor's pleasure. Fine for a Jira shop that already accepted that. Not mine, and I would not pretend otherwise.