Universal MCP Client

Connect to any MCP server without bloating context with tool definitions.

⚠️ PLAYWRIGHT USERS: READ "CRITICAL: Playwright Browser Session Behavior" SECTION BELOW!

Each MCP call = new browser session. Browser CLOSES after each call.
You CANNOT navigate in one call and click in another.
Use browserruncode for ANY multi-step operation.
If you need to return to a state (e.g., logged in), you MUST redo ALL steps from scratch.

How It Works

Instead of loading all MCP tool schemas into context, this client:

1. Lists available servers from config
2. Queries tool schemas on-demand
3. Executes tools with JSON arguments

Configuration

Config location priority:

1. MCPCONFIGPATH environment variable
2. .claude/skills/mcp-client/references/mcp-config.json
3. .mcp.json in current directory
4. ~/.claude.json

Commands

# List configured servers
python scripts/mcp_client.py servers

# List tools from a specific server
python scripts/mcp_client.py tools playwright

# Call a tool
python scripts/mcp_client.py call playwright browser_navigate '{"url": "https://example.com"}'

CRITICAL: Playwright Browser Session Behavior

⚠️ The Session Problem

Each MCP call creates a NEW browser session. The browser CLOSES after each call.

This means:

# ❌ WRONG - These run in SEPARATE browser sessions!
python scripts/mcp_client.py call playwright browser_navigate '{"url": "https://example.com"}'
python scripts/mcp_client.py call playwright browser_click '{"element": "Accept cookies"}'
python scripts/mcp_client.py call playwright browser_snapshot '{}'
# ^ The snapshot captures a FRESH page, not the page after clicking!

✅ The Solution: browserruncode

Use browserruncode to run multiple Playwright steps in ONE browser session:

python scripts/mcp_client.py call playwright browser_run_code '{
  "code": "
    await page.goto(\"https://example.com\");

    // Wait for and click cookie banner
    const acceptBtn = page.getByRole(\"button\", { name: /accept/i });
    if (await acceptBtn.isVisible({ timeout: 3000 }).catch(() => false)) {
      await acceptBtn.click();
      await page.waitForTimeout(1000);
    }

    // Wait for page to stabilize
    await page.waitForLoadState(\"networkidle\");

    // Return snapshot data for analysis
    const snapshot = await page.accessibility.snapshot();
    return JSON.stringify(snapshot, null, 2);
  "
}'

When to Use Each Approach

Playwright Workflows for Test Discovery

1. Basic Page Exploration (Single Step)

browser_navigate returns both navigation result AND accessibility snapshot:

python scripts/mcp_client.py call playwright browser_navigate '{"url": "https://example.com"}'

Output includes:

• Page URL and title
• Full accessibility tree (all visible elements with roles, names, states)
• Element references for further interaction

Use this when: Simple page load without interactions

2. Page with Cookie Banner (Multi-Step)

python scripts/mcp_client.py call playwright browser_run_code '{
  "code": "
    await page.goto(\"https://www.olx.ro\");

    // Handle cookie consent
    try {
      const cookieBtn = page.getByRole(\"button\", { name: \"Accept\" });
      await cookieBtn.click({ timeout: 5000 });
      await page.waitForTimeout(1000);
    } catch (e) {
      // No cookie banner
    }

    // Get accessibility snapshot
    const snapshot = await page.accessibility.snapshot({ interestingOnly: false });
    return JSON.stringify(snapshot, null, 2);
  "
}'

3. Navigate to Subpage (Multi-Step)

python scripts/mcp_client.py call playwright browser_run_code '{
  "code": "
    await page.goto(\"https://www.olx.ro\");

    // Dismiss cookies
    const acceptBtn = page.getByRole(\"button\", { name: \"Accept\" });
    if (await acceptBtn.isVisible({ timeout: 3000 }).catch(() => false)) {
      await acceptBtn.click();
      await page.waitForTimeout(500);
    }

    // Navigate to login
    await page.goto(\"https://www.olx.ro/cont/\");

    // Wait for redirect to login domain
    await page.waitForURL(/login\\.olx\\.ro/, { timeout: 10000 });

    // Get form structure
    const snapshot = await page.accessibility.snapshot();
    return JSON.stringify({ url: page.url(), snapshot }, null, 2);
  "
}'

4. Explore Element Interactions (Multi-Step)

Use this to understand how menus/dropdowns behave:

python scripts/mcp_client.py call playwright browser_run_code '{
  "code": "
    await page.goto(\"https://www.olx.ro\");

    // Dismiss cookies
    const acceptBtn = page.getByRole(\"button\", { name: \"Accept\" });
    if (await acceptBtn.isVisible({ timeout: 3000 }).catch(() => false)) {
      await acceptBtn.click();
    }

    // Click on category to see what happens
    const categoryLink = page.getByRole(\"link\", { name: /Auto, moto/i }).first();
    await categoryLink.click();

    // Wait to see result
    await page.waitForTimeout(1500);

    // Capture state after click
    const snapshot = await page.accessibility.snapshot();
    return JSON.stringify({
      url: page.url(),
      didNavigate: page.url().includes(\"auto\"),
      snapshot: snapshot
    }, null, 2);
  "
}'

5. Fill Form and Capture State

python scripts/mcp_client.py call playwright browser_run_code '{
  "code": "
    await page.goto(\"https://login.olx.ro\");

    // Fill login form
    await page.locator(\"input[type=email]\").fill(\"[email protected]\");
    await page.locator(\"input[type=password]\").fill(\"test123\");

    // Click login button
    await page.getByTestId(\"login-submit-button\").click();

    // Wait for response
    await page.waitForTimeout(3000);

    // Capture any error messages
    const errors = await page.locator(\"[class*=error], [role=alert]\").allTextContents();
    const snapshot = await page.accessibility.snapshot();

    return JSON.stringify({
      url: page.url(),
      errors: errors,
      snapshot: snapshot
    }, null, 2);
  "
}'

Gathering Selectors for Page Objects

Best Practices

1. Use Accessibility Tree First

The snapshot from browsernavigate or browserrun_code provides:

• Role: button, link, textbox, combobox, etc.
• Name: accessible name (from label, aria-label, text content)
• State: disabled, checked, expanded, etc.

Map these to Playwright locators:

// From snapshot: { role: "button", name: "Căutare" }
page.getByRole('button', { name: /Căutare/i })

// From snapshot: { role: "textbox", name: "Ce anume cauți?" }
page.getByRole('textbox', { name: /Ce anume cauți/i })

// From snapshot: { role: "link", name: "Auto, moto și ambarcațiuni" }
page.getByRole('link', { name: /Auto, moto/i })

2. Selector Priority

3. Handling Multiple Matches

// Use .first() when multiple match
page.getByRole('link', { name: 'Category' }).first()

// Use parent context
page.locator('nav').getByRole('link', { name: 'Category' })

// Use filter
page.getByRole('button').filter({ hasText: /submit/i })

4. Get Full DOM for Complex Cases

When accessibility tree isn't enough, get raw HTML:

python scripts/mcp_client.py call playwright browser_run_code '{
  "code": "
    await page.goto(\"https://example.com\");

    // Get specific element HTML
    const formHtml = await page.locator(\"form\").first().innerHTML();

    // Or get all buttons with their attributes
    const buttons = await page.locator(\"button\").evaluateAll(btns =>
      btns.map(b => ({
        text: b.textContent,
        testid: b.dataset.testid,
        class: b.className,
        type: b.type
      }))
    );

    return JSON.stringify({ formHtml, buttons }, null, 2);
  "
}'

Quick Reference: Playwright MCP Tools

Tool Arguments

browser_navigate

{"url": "https://example.com"}

browserruncode

{
  "code": "await page.goto('https://example.com'); return await page.title();"
}

The code must be valid JavaScript that:

• Uses page object (Playwright Page)
• Uses await for async operations
• Returns the data you want (use JSON.stringify for objects)

browser_click

{"element": "Submit button", "ref": "optional-element-ref"}

browser_type

{"element": "Email input", "text": "[email protected]"}

Error Handling

Setup

1. Copy the example config:

``bash cp .claude/skills/mcp-client/references/mcp-config.example.json \ .claude/skills/mcp-client/references/mcp-config.json ``

1. The config should contain:

``json { "mcpServers": { "playwright": { "command": "npx", "args": ["@playwright/mcp@latest"] } } } ``

1. Install dependencies:

``bash pip install mcp fastmcp ``

Config Example

See [references/mcp-config.example.json](references/mcp-config.example.json)

Available Servers

See [references/mcp-servers.md](references/mcp-servers.md) for:

• Playwright (browser automation)
• GitHub (repository operations)
• Filesystem (file access)
• Sequential Thinking (reasoning)
• And more...

Dependencies

pip install mcp fastmcp