AutoDev MCP (Dual-Track Testing)

AutoDev MCP is a TypeScript MCP server that keeps your original CLI testing and adds an extra Chrome Dev (browser) testing track with persistent logs.

What it does

• Keeps original test flow via CLI (run_tests) - not replaced

• Adds browser test flow (browser_* tools) via Playwright + CDP events

• Persists everything to SQLite:

  • sessions

  • logs (cli/browser/console/error/network/assertion)

  • screenshots

• Supports combined execution via run_pipeline

• After pipeline completion, automatically provides 10 requirement/optimization directions (中文输出)

Install

npm install
npm run build

Run MCP server (stdio)

npm run start

For local development:

npm run dev

Optional config: autodev.config.json

{
  "projectRoot": ".",
  "storage": {
    "databasePath": ".autodev/logs/autodev.db",
    "screenshotsDir": ".autodev/logs/screenshots"
  },
  "commands": {
    "test": "npm test"
  },
  "browser": {
    "headless": true,
    "timeoutMs": 15000
  },
  "pipeline": {
    "maxIterations": 3
  },
  "compact": {
    "enabled": true,
    "warningThreshold": 0.8,
    "compactThreshold": 1.0,
    "estimatedMaxChars": 120000
  },
  "git": {
    "enabled": true,
    "autoCommitOnFeatureComplete": true,
    "defaultCommitPrefix": "feat",
    "protectSensitiveFiles": true,
    "sensitiveFilePatterns": [
      "(^|/)\\.env(\\..*)?$",
      "(^|/)credentials\\.json$",
      "(^|/)secret(s)?(\\.|$)",
      "(^|/).+\\.pem$",
      "(^|/).+\\.key$"
    ]
  }
}

Tool overview

CLI track (original)

• run_tests

• generate_tests (stub)

• fix_failures (stub)

Browser track (new, additional)

• browser_open

• browser_navigate

• browser_interact

• browser_assert

• browser_screenshot

• browser_get_console

• browser_get_errors

• browser_get_network

• browser_close

Orchestration/query

• run_pipeline (CLI first, browser optional)

• propose_requirements (always outputs 10 directions)

• check_compact (检查上下文占用并给出 none/warn/compact 建议)

• git_commit_feature（小功能完成后按策略提交 git）

• git_status_summary（查看当前仓库改动状态）

• test_get_session_logs

• test_compare_sessions

Example workflow

1. Run original tests:

   • call run_tests

2. Start browser session:

   • call browser_open with URL

3. Interact and assert:

   • browser_interact

   • browser_assert

4. Capture screenshot if needed:

   • browser_screenshot

5. Close and finalize summary:

   • browser_close

6. Query persisted logs:

   • test_get_session_logs

7. Generate next-cycle demand directions:

   • propose_requirements

run_pipeline also includes suggestedRequirements in the final output so you can directly pick your next approved requirement.

run_pipeline now also returns compactSuggestion:

• none: 上下文充足，继续执行

• warn: 上下文超过预警阈值（默认 80%），建议当前小任务结束后 compact

• compact: 上下文达到 compact 阈值（默认 100%），建议立即 compact

run_pipeline 现在也会在测试通过后尝试执行 gitCommit：

• 若目录是 git 仓库且存在可提交改动，会自动提交

• 若不是 git 仓库，自动跳过并给出原因

• 默认会拦截疑似敏感文件（如 .env、*.pem、*.key）

测试日志与截图位置

• SQLite 数据库：.autodev/logs/autodev.db

• 截图目录：.autodev/logs/screenshots

你可以通过 MCP 工具查看：

• test_get_session_logs：按 sessionId 拉取日志、摘要、截图路径

• test_compare_sessions：对比两次测试结果

评分标准（需求方向）

每条方向的评分由以下字段组成：

• impact：影响面（1-10）

• confidence：把握度（1-10）

• effort：工作量映射分值（S=2, M=5, L=8）

公式：

priorityScore = (impact × confidence) / effort

分数越高，优先级越高。

Notes

• Browser track is explicitly additive and independent from original CLI tests.

• generate_tests and fix_failures are stubs in this version and intentionally logged for future extension.

• Requirement generation is evidence-based (sessions/logs/assertions/errors) and returns exactly 10 scored directions each cycle (中文输出).