pr-review-mcp

English | Русский

pr-review-mcp

Unified MCP control plane for AI pull request reviews on GitHub.

Modern pull requests often attract multiple AI reviewers, but their comments arrive in different formats, different places, and different timelines. pr-review-mcp turns that noise into one MCP-native workflow: it normalizes review output from seven agent sources, exposes 19 focused tools, and adds orchestration primitives for parallel review handling.

If you use Claude Code, Claude Desktop, or another MCP client to process GitHub reviews, this server gives you one place to list findings, inspect details, invoke agents, wait for them server-side, and coordinate worker agents without building your own review pipeline.

What's New in v0.3.0

• pr_await_reviews adds a server-side wait loop for AI reviews, so clients no longer need to poll GitHub manually while agents finish posting.
• pr_invoke now returns since, invokedAgentIds, and awaitHint, making the handoff into pr_await_reviews explicit and reliable.
• Claude Code integration is stronger with skills/review/SKILL.md, a dedicated skill that wraps /pr:review into an autonomous review workflow.
• CI now runs build, tests, and coverage on Node 20 and 22 through GitHub Actions.
• Coverage reporting is part of the release baseline with @vitest/coverage-v8; v0.3.0 records 225 Vitest test cases and a 51.7% statement coverage baseline.
• Qodo completion tracking is more accurate because persistent issue comments are detected via updated_at, not only created_at.
• Agent status polling was simplified by extracting shared logic into fetchAgentStatusForAgents, removing duplicated code paths.
• Ten dependency vulnerabilities were resolved as part of the release hardening work.

Features

• Automates multi-agent PR review workflows with 19 MCP tools, 3 prompts, and 1 dynamic PR resource.
• Eliminates review-source fragmentation by normalizing comments from CodeRabbit, Gemini, Copilot, Sourcery, Qodo, Codex, and Greptile.
• Prevents polling loops in clients with pr_await_reviews, a server-side monitor for agent completion.
• Coordinates parallel workers safely with file-level partition claiming, progress reporting, and orchestration status checks.
• Surfaces machine-readable review data through structured outputs on key tools such as pr_summary, pr_list, pr_get, pr_get_work_status, and pr_progress_check.
• Supports local and shared deployments through stdio by default and StreamableHTTP with pr-review-mcp --http or pr-review-mcp --http 8080.
• Protects operational workflows with confirmation flows for destructive actions such as merge and coordination reset.
• Ships with Claude Code assets including .claude-plugin/plugin.json, .mcp.json, and the review skill for slash-command driven orchestration.

Architecture

graph TD
    Client["MCP Client<br/>Claude Code / Claude Desktop / Inspector"]

    Client -->|stdio| Server
    Client -->|StreamableHTTP| Server

    subgraph Server["pr-review-mcp v0.3.0"]
        subgraph Interface["MCP Interface"]
            Resource["Resource<br/>pr://{owner}/{repo}/{pr}"]
            PromptReview["Prompt<br/>review"]
            PromptReviewBg["Prompt<br/>review-background"]
            PromptSetup["Prompt<br/>setup"]
        end

        subgraph Analysis["Analysis Tools (7)"]
            T1["pr_summary"]
            T2["pr_list_prs"]
            T3["pr_list"]
            T4["pr_get"]
            T5["pr_changes"]
            T6["pr_poll_updates"]
            T7["pr_await_reviews"]
        end

        subgraph Action["Action Tools (6)"]
            T8["pr_invoke"]
            T9["pr_resolve"]
            T10["pr_labels"]
            T11["pr_reviewers"]
            T12["pr_create"]
            T13["pr_merge"]
        end