Dejank

Give your AI coding agent hands on a live React Native app.

Dejank is an MCP server. Your agent can tap the UI, read the screen, pull native logs, capture React/JS profiles, and save evidence — so it can find bugs, verify fixes, and investigate performance without you switching to DevTools by hand.

You describe what's wrong in plain English. The agent uses Dejank to look at the running app and figure it out.

What you need first

• A React Native or Expo app running in debug mode
• Metro already started (usually port 8081)
• A simulator or device with the app open
• An AI tool that supports MCP (Cursor, Claude, VS Code, ChatGPT desktop, Windsurf, Antigravity, Gemini CLI, …)

Dejank doesn't start Metro or your app for you — it attaches to what's already running.

Install

1. Add Dejank to your AI tool

From this repo:

pnpm mcp:install-global -- --npx

That writes a dejank MCP server into every client it finds on your machine:

| App | Config | |---|---| | Cursor | ~/.cursor/mcp.json | | Claude Desktop | ~/Library/Application Support/Claude/claude_desktop_config.json | | Claude Code | ~/.claude.json (if claude CLI is installed) | | VS Code / VSCodium | …/Code/User/mcp.json (servers key) | | ChatGPT desktop | ~/.codex/config.toml + …/ChatGPT/chatgpt_mcp_config.json | | Windsurf | ~/.codeium/windsurf/mcp_config.json | | Antigravity | ~/.gemini/antigravity/mcp_config.json | | Gemini CLI | ~/.gemini/settings.json | | GitHub Copilot (Agent Host) | ~/.copilot/mcp-config.json | | Warp | ~/.warp/.mcp.json | | Amazon Q / Kiro / LM Studio / Cline / Roo | when their config dirs already exist |

Equivalent Cursor / Claude config:

{
  "mcpServers": {
    "dejank": {
      "command": "npx",
      "args": ["-y", "dejank"],
      "env": {
        "DEJANK_METRO_HOST": "localhost",
        "DEJANK_METRO_PORT": "8081"
      }
    }
  }
}

VS Code uses the same command, but the top-level key is servers (not mcpServers). ChatGPT desktop on macOS stores MCP in ~/.codex/config.toml under [mcp_servers.dejank].

Developing Dejank itself? Use the local build instead of npm:

pnpm mcp:install-global

Fully quit and reopen your AI tools so they reload MCP. In VS Code, use Agent mode in Copilot Chat.

2. Start your app

npx expo start
# open the app on a simulator or device

Metro should be running (usually on port 8081) before you ask the agent anything.

Try it

Just ask your agent, in plain English:

Reproduce the bug when I tap Submit and tell me what failed.

I fixed the null check — tap Submit again and confirm the error is gone.

Why is the screen slow when I press Load feed?

Show me the UI tree and a screenshot of the current screen.

Pull the last minute of device logs after I tap Retry.

The agent picks the right Dejank tools for the job — no need to know their names.

What it's good for

| Goal | Dejank gives the agent | |---|---| | Fix a bug | UI tree, tap, screenshot, native logs, component tree | | Verify a fix | The same tap/screenshot/logs, repeatable before vs after | | Diagnose slowness | One-shot JS + React profiling around a tap, via investigate_slow_screen | | Check UI state | UI tree + screenshot | | Trace React renders | Component tree and profiler data via Fusebox DevTools | | Catch runtime noise | Android logcat / iOS Simulator logs |

Dejank collects evidence from the live app. Your agent — and you — decide the fix.

Not built yet: dedicated JS console streaming, network inspection, memory/heap profiling. Native device logs often catch JS/runtime errors anyway, since most get printed there too.

How it works

you ask to debug / verify / profile
            │
            ▼
   agent calls Dejank (MCP)
            │
            ▼
 Dejank attaches to Metro / the app
            │
            ▼
tap · screenshot · logs · JS/React profiles
            │
            ▼
   evidence returns to the agent
            │
            ▼
 agent explains it, and can patch your code

Tools

| Tool | Use it for | |---|---| | investigate_slow_screen | Performance — one-shot JS + React profiling + diagnosis around an optional tap | | get_ui_tree / tap | See on-screen controls; reproduce taps by label or coordinates | | take_screenshot | Visual proof for bugs and UI checks | | get_device_logs | Native logs (logcat / iOS Simulator) | | get_component_tree | React component names via Fusebox DevTools | | capture_* / start_* / stop_* profiles | Finer-grained JS or React profiling when you need more control | | list_targets / get_target_info | Metro debug targets, useful if attaching fails | | list_artifacts | Find recent evidence files |

For "is it slow?" questions, investigate_slow_screen alone is usually enough. For bugs and validation, the agent composes tap + get_ui_tree + take_screenshot + get_device_logs.

Settings

| Env var | Meaning | Default | |---|---|---| | DEJANK_METRO_HOST | Where Metro runs | localhost | | DEJANK_METRO_PORT | Metro port | 8081 | | DEJANK_ARTIFACT_DIR | Where profiles and screenshots are saved | .dejank/artifacts |

Developing Dejank itself

pnpm install
pnpm build
pnpm mcp:install-global   # local dist/ → Cursor / Claude (use -- --npx for npm)
pnpm mcp                  # run the server in the foreground

Contributor checks:

pnpm validate:m0   # basic attach / capability check
pnpm validate:m3   # injected slow-screen scenarios (sample app + simulator)

Historical product notes: docs/SCOPE.md.

License

MIT