playwright-mcp vs code-review-graph
Two of the top projects in MCP servers for developers, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
playwright-mcp
MCP server that lets LLMs drive a browser through Playwright accessibility snapshots
code-review-graph
Local Tree-sitter code graph that gives coding assistants review context over MCP
Verdict
Both score 92; playwright-mcp ranks higher on stars, 39,500 to 32,716. Of the 4 score parts we measure for both, playwright-mcp leads on 2, code-review-graph on 1 and 1 is even.
- Adoptionplaywright-mcp leads by 3 points, 93 to 90: close to even.
- FreshnessEven at 100.
- Maintenanceplaywright-mcp leads by 7 points, 100 to 93.
- Agent-readycode-review-graph leads by 25 points, 70 to 45: a wide gap.
Pick playwright-mcp if you are
Developers giving LLM agents browser automation through MCP clients
- Uses accessibility snapshots, so text-only models can operate pages without vision
- Works with many MCP clients: VS Code, Cursor, Claude Code, Codex, Gemini CLI, Goose and others
Pick code-review-graph if you are
Developers who want coding assistants to read fewer files when reviewing changes
- One install command detects and configures about 16 AI coding tools, with a clean uninstall
- Incremental updates re-parse only SHA-256-changed files; two-file edit on django took about 2.5 s
| What we compare | playwright-mcp | code-review-graph |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 93, widely used | 90, widely used |
| Momentum | Not scored for this kind | Not scored for this kind |
| Freshness | 100, active | 100, active |
| Maintenance | 100, healthy | 93, healthy |
| Agent-ready | 45, minimal | 70, partly |
| Facts from GitHub and the README | ||
| Stars | 39.5k | 32.7k |
| License | Apache-2.0 (permissive) | MIT (permissive) |
| Last commit | Nov 2026 | Nov 2026 |
| Last release | Nov 2026 | Nov 2026 |
| Language | TypeScript | Python |
| Docker | Yes | Yes |
| GPU | Not needed | Not needed |
| arm64 or Apple Silicon | Not stated | Not stated |
playwright-mcp
Runs as an MCP server, started with `npx @playwright/mcp@latest`, and exposes Playwright browser control to any MCP client. It feeds the model structured accessibility-tree snapshots instead of screenshots, so no vision model is required. Options cover Chrome, Firefox, WebKit and Edge, CDP or extension attachment to a running browser, proxies, device emulation and origin allow/block lists.
Who it is for: Developers giving LLM agents browser automation through MCP clients
Strengths
- Uses accessibility snapshots, so text-only models can operate pages without vision
- Works with many MCP clients: VS Code, Cursor, Claude Code, Codex, Gemini CLI, Goose and others
- Can attach to a running Chrome or Edge via CDP endpoint or the Playwright Extension
- Many flags: isolated profiles, secrets file, URL param redaction, init scripts, proxy, device emulation
Weaknesses
- README says tool schemas and verbose snapshots cost more tokens than the Playwright CLI with skills
- README recommends the separate Playwright CLI for high-throughput coding agents
- Origin allow/block lists are documented as not a security boundary
- Headed browser by default, so servers need --headless
- no GPU
- Docker
- Needs Node.js 18 or newer, Playwright browsers (Chrome, Firefox, WebKit or Edge), MCP client
- Models: Any LLM via an MCP client
code-review-graph
Parses a repository with Tree-sitter into a SQLite graph of functions, classes, imports, calls and tests, then serves blast-radius and risk-scored context over MCP so the assistant reads only affected files. The installer configures about 16 tools (Claude Code, Cursor, Codex, Windsurf, Zed and others) and updates incrementally via hooks, git hooks or watch mode. A GitHub Action posts risk-scored PR comments.
Who it is for: Developers who want coding assistants to read fewer files when reviewing changes
Strengths
- One install command detects and configures about 16 AI coding tools, with a clean uninstall
- Incremental updates re-parse only SHA-256-changed files; two-file edit on django took about 2.5 s
- Covers 30+ languages plus notebooks, and custom grammars via languages.toml
- Runs locally; the GitHub Action builds the graph on the CI runner without sending code out
Weaknesses
- Impact recall of 1.0 is circular, since ground truth comes from the same graph
- Against a grep baseline, median token reduction is 6.65x, not the 63x headline
- Graph context can cost more tokens than a plain read for small single-file edits
- Cold build of a roughly 3,000-file repository took about 40 seconds
- no GPU
- Docker
- Needs Python 3.10+, Tree-sitter, SQLite
- Models: all-MiniLM-L6-v2 (local embeddings)
More in MCP servers for developers
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.