Bugbotが学習ルールとMCP対応を獲得 Bugbot Learned Rules and MCP Support

Cursor has rolled out two significant upgrades to Bugbot, its AI-powered code review tool: an automatic rule-learning mechanism that derives review criteria from past feedback, and integration support for Model Context Protocol (MCP) servers. Together, the changes are aimed at making Bugbot's comments more attuned to the specific conventions and context of each engineering organization.

The learned-rules feature works by observing how developers respond to Bugbot's suggestions over time. When a team repeatedly dismisses certain types of comments, applies similar fixes, or pushes back on particular kinds of warnings, Bugbot accumulates those patterns and reuses them as implicit review standards. The intended outcome is fewer repeated false positives and less need to spell out coding conventions on every pull request. In effect, team-specific habits and review culture should propagate into Bugbot's behavior without anyone having to author explicit rule files, although Cursor still allows manually defined rules for cases where teams want tighter control.

The second addition, MCP support, lets Bugbot pull in information from external data sources and tools while performing a review. MCP is an open specification originally proposed by Anthropic that standardizes how AI agents connect to outside systems. By pointing Bugbot at MCP servers for issue trackers, internal documentation, design specs, or database schemas, teams can in principle ask the reviewer to check whether a code change actually satisfies the linked ticket's requirements, whether it conflicts with an existing schema, or whether it contradicts documented architectural decisions. That moves Bugbot beyond the boundaries of the diff itself and into the surrounding organizational context that human reviewers typically rely on.

The two features are complementary. Learned rules capture tacit knowledge that emerges from review interactions, while MCP exposes explicit knowledge stored elsewhere in a company's stack. Used together, they appear designed to address one of the most common complaints about generic AI code review: that suggestions are technically reasonable but disconnected from how a specific codebase or team actually operates.

AI code review has become a crowded space. GitHub's Copilot Code Review, Graphite's Diamond, CodeRabbit, Greptile, and several smaller entrants are all competing to embed AI reviewers into pull request workflows. As the underlying models converge in raw capability, differentiation is increasingly shifting toward how well a tool can ingest organization-specific context — repository history, prior reviews, internal documentation, ticketing systems — rather than how well it performs generic static analysis. Cursor's latest update fits squarely within that trend.

There are also open questions that the announcement does not fully resolve. Learned rules raise the usual concerns about transparency and drift: if Bugbot internalizes patterns from past behavior, teams may want visibility into what it has inferred and the ability to audit or override those inferences, particularly when reviewers' historical responses were inconsistent. MCP integration similarly broadens the surface area for both capability and risk, since granting a reviewer agent access to internal systems implies trust boundaries that organizations will need to manage carefully. Cursor has not detailed how learned rules can be inspected or edited in the changelog entry, and behavior in practice will likely become clearer as teams adopt the features.

For existing Bugbot users, the practical upshot is that reviews should gradually feel more aligned with house style, and that integrators can begin wiring in the same MCP servers they may already be using elsewhere in the Cursor ecosystem. For the broader market, the update is another signal that the next phase of AI-assisted code review will be defined less by model quality alone and more by how deeply these tools can be embedded into the context where engineering work actually happens.