Before you start#
You need Node 22.19 or newer (Node 22 LTS recommended), access to a deployed Rovara project, and your installation’s control-plane URL from your admin. Developers do not need AWS credentials. Install an environment first →
The package is @preplabsai/rovara-code. The MCP server remains agentx, and tools retain the agentx_ prefix.
1. Sign in#
npx @preplabsai/rovara-code login <your-control-plane-url>
Sign in through the browser with Slack or your company’s configured sign-in. The CLI shows your identity and available projects. Use the URL your administrator gives you; it is your team’s deployment, not a Rovara-hosted service.
2. Add the MCP connection#
Run the command for your client, as yourself (without sudo):
Claude Code
claude mcp add --scope user agentx -- npx -y @preplabsai/rovara-code mcp
Or use npx @preplabsai/rovara-code mcp install --client claude-code.
Codex
npx @preplabsai/rovara-code mcp install --client codex
Cursor
npx @preplabsai/rovara-code mcp install --client cursor
Reopen your client if needed. mcp install pins the CLI version in the client entry; rerun it to update. Add --env <name> before mcp install if your administrator supplied an environment name.
3. Start a task#
After connecting, ask your client to list your Rovara projects and start a task in the chosen project, for example “Hand Rovara the navigation fix in <project-name> and run the relevant tests.” This direct-task path uses agentx_start_task and has no plan approvals. It is available in v0.3.4.
Start a Quick or Full workflow#
Rovara v0.3.4 includes the native workflow. This page follows the inspected mainline source, which may include later changes. Deployment settings and feature flags still apply. Release and coverage details →
Ask your client to find your projects and start a workflow. workflow_path is required: quick asks once for a coding plan; full asks for requirements, design and the coding plan.
List my Rovara Code projects.
Start a Quick workflow in <project-name> to improve the docs Copy button.
Reset its label after copying, announce success accessibly,
and preserve manual copy on clipboard failure.
Show me the plan before implementing.
The client uses agentx_start_workflow. Include the context the remote task needs; it does not receive your local conversation.
Read and decide on the plan#
agentx_get_task shows the current approval document, its digest and the workflow revision. Ask the client to show the whole document before you decide, or open the task page.
Your explicit decision calls agentx_decide_workflow with that revision and digest. An assistant should not approve a plan on your behalf without your instruction.
Show the current plan for task <task-id>, including checks and scope.
Follow checks, reviews and the draft PR#
After approval, checks, reviews and the draft PR follow on their own. A wait can time out while work continues; get the current task before retrying. When a step blocks, the owner chooses:
-
agentx_send_workflow_back: give the failed checks and findings the change introduced to coding. -
agentx_resolve_review_findings: accept each blocking finding with a reason, or have it fixed. -
agentx_review_workflow_candidate: retry a blocked review on the same code. -
agentx_retry_workfloworagentx_retry_workflow_publish: restart blocked planning, coding or checks, or retry opening the PR.
For PR comments, read the proposal with agentx_get_pr_feedback and decide with agentx_decide_pr_feedback. Tool reference →
Use your own ChatGPT plan#
If your admin turned on Sign in with ChatGPT, connect your Plus or Pro plan so your tasks run on it:
npx @preplabsai/rovara-code chatgpt connect