Cubic

What the cubic MCP server does

The cubic MCP server gives AI agents access to cubic's hosted review data for GitHub repositories. It supports reading open pull request findings with their file locations, severity, and review state, then starting a cloud review when a pull request needs a fresh analysis. Agents can also resolve, reopen, or dismiss findings; those changes are recorded on the related GitHub review thread.

Beyond pull requests, the cubic MCP server exposes repository wikis and individual pages, codebase scan results, scan issue details, and team review learnings. Organization-level access includes subscription and member information. Depending on account permissions and subscription, organization administrators may also manage seats, roles, and supported billing settings.

How it works

cubic hosts the endpoint at https://www.cubic.dev/api/mcp using Streamable HTTP. No local process or repository installation is required. An MCP client connects to that URL and authenticates either through the browser-based OAuth flow or with a personal API key sent as a bearer token.

The available operations follow the permissions of the authenticated cubic account. Connecting the endpoint does not grant access to repositories or organizations that the account cannot already use. Personal API keys also retain the permissions of the user who created them.

Pull request inspection and review execution are separate actions. get_pr_issues reads findings and review status; it does not launch a review. Use trigger_pr_review to request a cloud review. Pull request issue identifiers returned by get_pr_issues are used with update_pr_issue_status, while scan issue identifiers from get_scan are used with get_issue or update_issue_status.

Setup and configuration

Add the hosted URL to an MCP client that supports HTTP servers. Cursor can use ~/.cursor/mcp.json for all projects or .cursor/mcp.json for one project. Claude Code supports claude mcp add --transport http, and Codex CLI supports codex mcp add, followed by codex mcp login. In VS Code, use the MCP server command and select HTTP, enter the endpoint, name it cubic, and complete OAuth in the browser.

For clients or automation without OAuth, create a personal API key in cubic's MCP integration settings and provide it as an Authorization: Bearer header. Store keys in a secret manager for scripts and CI. In VS Code extension-host sessions, an input prompt can keep the key out of workspace configuration.

Tools and capabilities

The cubic MCP server supports workflows such as:

• Listing open findings for an accessible pull request.
• Starting a cloud review for a pull request.
• Changing a pull request finding to resolved, reopened, or dismissed.
• Listing wiki pages and reading a selected page.
• Browsing scan results, inspecting issues, and changing scan triage status.
• Reading review learnings for a repository.
• Inspecting organization subscriptions and members, with additional administrative controls where permitted.

Limitations and notes

An empty pull request findings result only indicates that the current commit has no open findings when the review has completed and no newer commits are present. Use the review trigger when those conditions are not met. The hosted MCP endpoint is for cubic's remote workflows; local code reviews instead use the cubic CLI. Tool availability depends on the account's permissions and subscription, and API keys are personal rather than shared organization credentials.