react-grab（React grab 前端）

react-grab — Browser element context for AI agents

Keyword: react-grab · grab · element context Point at a live React UI element, capture the exact component/file/HTML context, and feed it straight into Claude Code, Cursor, Copilot, Codex, Gemini, or an MCP workflow.

When to use this skill

• Install react-grab in a React project so agents can copy exact component context from the browser
• Choose between clipboard flow, MCP flow, and plugin/API flow for React element capture
• Fix react-grab setup issues like wrong load point, missing file paths, or empty MCP responses
• Add or remove an agent integration (claude-code, cursor, codex, gemini, mcp, etc.)
• Build a custom plugin or use the primitives API for a project-specific workflow

Do not use this skill for:

• Fresh-session browser automation, auth flows, or disposable browser verification → agent-browser
• Running-browser reuse with existing cookies/extensions/logins → playwriter
• Human feedback / annotation packets on rendered UI → agentation
• React performance diagnosis, rerender churn, hydration, bundle, or RSC/client-boundary work → react-best-practices
• General UI-system / token / component-library design work → design-system

Instructions

Step 1: Classify the request

Choose one primary mode before giving commands:

1. Clipboard install/use — user wants to hover, press Cmd/Ctrl+C, and paste context into an AI agent.
2. MCP workflow — user wants the agent to call a tool like get_element_context after browser selection.
3. Plugin/API workflow — user wants custom actions, lifecycle hooks, formatting, or primitives like freeze().
4. Troubleshooting/removal — user wants to fix setup problems or remove the integration cleanly.

Step 2: Confirm prerequisites truthfully

Use the upstream project as source of truth:

• npx grab@latest init is the default install path.
• Upstream package.json currently requires Node.js >=22.
• The browser overlay is a development-only tool; never present it as a production feature.

If the user’s environment is older than Node 22, say so explicitly before giving the install path.

Step 3: Give the shortest correct path

Mode A — Clipboard install/use

Use this as the default path unless the user explicitly asks for MCP or plugins:

# At the project root
npx -y grab@latest init

Then explain the runtime flow:

1. Start the dev server.
2. Open the app in the browser.
3. Hover the target element.
4. Press Cmd+C (Mac) or Ctrl+C (Windows/Linux).
5. Paste the captured context into the AI agent.

Clipboard output should include:

• component name
• source file path
• line/column when available
• HTML snippet

If the user needs framework-specific setup, use references/install-and-routing.md.

Mode B — MCP workflow

When the user wants programmatic access from the agent instead of manual paste:

npx -y grab@latest add mcp

Explain the real flow clearly:

1. User selects or hovers the element in the browser overlay.
2. The MCP server exposes the latest selection.
3. The AI agent calls get_element_context.
4. The tool returns file path, component, stack, and HTML context.

If the user asks how this differs from clipboard flow:

• Clipboard is manual paste into chat.
• MCP is tool-based retrieval after selection.

Use references/agent-workflows.md for agent-specific install and routing guidance.

Mode C — Plugin/API workflow

When the user wants custom behavior, automation, or a deeper integration:

• Route installation and workflow choice first.
• Then point to:

- references/api.md for full API surface - references/agent-workflows.md for practical plugin/MCP patterns

Common primitives to mention:

• getElementContext(element)
• freeze() / unfreeze()
• openFile(filePath, lineNumber?)
• registerPlugin(...)

Do not dump the full API unless the user asked for the deep path.

Mode D — Troubleshooting/removal

If the user wants fixes or teardown:

• Use references/troubleshooting.md for symptom-first triage.
• For removal, use upstream CLI removal commands:

npx -y grab@latest remove <agent>

Step 4: Keep the boundary sharp

Keep react-grab focused on React element context capture from the browser.

Route out when the user is actually asking for:

• general browser automation → agent-browser or playwriter
• UI review / visual annotation → agentation
• performance / profiling / hydration / bundle / RSC issues → react-best-practices
• reusable component-system design → design-system or ui-component-patterns

Step 5: Prefer reference docs over bloating the front door

After classifying the mode, load only the matching support doc:

• references/install-and-routing.md
• references/agent-workflows.md
• references/troubleshooting.md
• references/api.md

Keep the main answer short unless the user explicitly asks for the long path.

Examples

Example 1: install in a Next.js app and use clipboard flow

User: “Set up react-grab in my Next.js app so I can copy a component into Claude Code.”

Response shape:

1. Confirm Node 22+ prerequisite.
2. Use npx -y grab@latest init as the default path.
3. Explain dev-server → browser hover → Cmd/Ctrl+C → paste into Claude Code.
4. Only load framework-specific setup if auto-detect is not viable.

Example 2: MCP setup for agent tool access

User: “Install react-grab MCP so the agent can grab the selected button context.”

Response shape:

1. Confirm Node 22+ prerequisite.
2. Run npx -y grab@latest add mcp.
3. Explain that the browser selection happens first, then the agent calls get_element_context.
4. Route generic browser-control questions to agent-browser / playwriter if needed.

Example 3: custom plugin workflow

User: “Create a react-grab plugin that adds a Jira action to the context menu.”

Response shape:

1. Keep react-grab as the plugin/API lane.
2. Use registerPlugin(...) with an action and optional hooks.
3. Point to references/api.md for the full hook surface.
4. Keep the example focused on the requested action rather than reproducing the whole API.

Example 4: fix setup drift

User: “grab init keeps failing on my machine.”

Response shape:

1. Check Node version first and call out the Node 22+ requirement.
2. Confirm they are running from the project root.
3. Use references/troubleshooting.md for overlay/load/MCP/path failures.
4. Route unrelated framework build issues to the appropriate frontend skill.

Best practices

1. State the Node.js >=22 prerequisite truthfully before recommending install commands.
2. Default to npx -y grab@latest init unless the user explicitly needs manual framework setup.
3. Keep it dev-only — never imply this belongs in production builds.
4. Explain clipboard vs MCP clearly so users do not confuse paste-based and tool-based workflows.
5. Route broad browser or UI-review work out early instead of letting react-grab absorb adjacent skills.
6. Use support docs for depth instead of pasting long installation or API manuals into the main response.

References

• react-grab GitHub
• react-grab.com
• Install and routing guide
• Agent workflows
• Troubleshooting
• Full API reference
• Helper scripts — install.sh and add-agent.sh