UI Debugger MCP

An MCP server that debugs UIs autonomously — so the AI that wrote your app can also test it, without a human clicking through every flow.

The problem

AI coding agents (Claude, etc.) are great at writing code. They're bad at knowing if the UI actually works. For backend code there are unit and integration tests. For UI, a human still has to open the app, log in, click around, and report what's broken. That human-in-the-loop is slow, boring, and the main bottleneck when an entire product is built by AI.

Related MCP server: Curupira

The idea

Eliminate the human from the UI-debug loop with an MCP server.

• A smart agent (Claude Code, Cursor, …) finishes a PR and wants to verify the UI.

• It hands a story to this server: "on web, log in and do X, Y, Z — tell me if it breaks."

• A small fast agent runs inside this server (via the Vercel AI SDK). It drives the browser or desktop, watches console + network, takes screenshots.

• It reports structured findings back: pass/fail, what broke, evidence.

• The smart agent fixes the code and asks again. Loop until the UI works.

Unlike playwright-mcp — where the smart model issues every single click itself — here the smart model stays high-level and delegates the whole clicking loop to the small agent.

How it's different from playwright-mcp

Architecture — the three actors

Picture a boss, a fast blind driver, and a describer with eyes:

   ┌─────────────┐   MCP conversation    ┌──────────────────────────────────────┐
   │ smart agent │  start_debug ───────▶ │        UI Debugger MCP server         │
   │  (Claude)   │  send_message (live)  │                                       │
   │             │ ◀─────── get_findings │   ┌────────────┐     ┌────────────┐   │
   │ sets goals  │                       │   │  fast guy  │ look│ vision guy │   │
   │ fixes code  │                       │   │  (driver)  │────▶│  (eyes)    │   │
   │ loops       │                       │   │ deepseek   │◀────│  glm 5v    │   │
   └─────────────┘                       │   │ text·blind │ desc│ image      │   │
          ▲                              │   └─────┬──────┘     └────────────┘   │
          │ "works + looks nice"         │     observe / act (SQL-like)          │
          │ findings + screenshots       │         │ shared adapter contract     │
          └──────────────────────────────│─────────┼─────────────────────────────│
                                          └─────────┼─────────────────────────────┘
                                                    ▼
                              ┌──────────────┬──────────────┬──────────────┐
                              │  web (CDP)   │ desktop      │ android      │
                              │  browser     │ X11/Wayland  │ ADB          │
                              └──────────────┴──────────────┴──────────────┘

• smart agent — the boss (Claude/caller). Sends a goal, reads findings, fixes the code, loops. Stays high-level — never clicks.

• fast guy — the driver. Fast, cheap, text-only and blind. Runs the click loop on structure (DOM / a11y tree / view hierarchy). Default: deepseek.

• vision guy — the eyes. Multimodal. The driver calls look to ask "does this look right? is the button centred?" and gets a description back. Default: glm. Spent only when visual judgment is needed.

One goal: the UI works and looks nice. Full design in docs/idea/.

Every run keeps its screenshots and stitches them into a short captioned replay video — Claude attaches it to the PR so a reviewer sees the flow working in ~10 seconds (docs/idea/workspace.md).

Targets

One project can expose several debug targets. A large app can have all three:

Three adapters, one shared contract. Each runs managed (server launches the target) or attach (connect to a running one via cdpUrl / adbSerial). Linux first. iOS is out of scope on Linux (macOS-only tooling).

A managed android run boots the emulator windowless — -no-window with -gpu swiftshader_indirect -no-audio, because -no-window alone still opens a host GL context and dies on a box with no display. Set "headless": false on the target to watch it in a window; append anything else with emulatorArgs.

Setup

Install like any local MCP server — one entry in your .mcp.json:

{
  "mcpServers": {
    "ui-debugger": {
      "command": "npx",
      "args": ["-y", "@developerz.ai/ui-debugger-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-...",
        "OPENAI_BASE_URL": "https://openrouter.ai/api/v1"
      }
    }
  }
}

Wiring more than one repo? .mcp.json is usually committed, so don't paste the key into each one — export it once in your shell and reference it, which keeps every repo's entry identical and secret-free:

"env": {
  "OPENAI_API_KEY": "${UI_DEBUGGER_API_KEY}",
  "OPENAI_BASE_URL": "${UI_DEBUGGER_BASE_URL:-https://openrouter.ai/api/v1}"
}

It's also published in the official MCP Registry as io.github.developerz-ai/ui-debugger-mcp — any client that browses the registry (instead of a hand-written .mcp.json entry) can find and install it by that name.

Then add a per-project .ui-debugger-mcp.json describing the app to debug (models, targets, urls). The fastest way is the init command:

npx @developerz.ai/ui-debugger-mcp@latest init   # in your project root

ui-debugger-mcp init scaffolds a project for debugging (described in docs/idea/config.md):

• creates the workspace dir ./tmp/ui-debugger-mcp/

• writes a starter .ui-debugger-mcp.json (default deepseek/glm models, a web target stub) if one doesn't already exist

• adds tmp/ to .gitignore

• prints the .mcp.json snippet to paste (it never writes your API key)

Config files:

• .mcp.json → how to launch the server (command + secret key). Gitignored.

• .ui-debugger-mcp.json → how to debug this app (models, targets). Committed.

Every key of both files is documented in docs/idea/config.md; every tool's exact input/output shape is in docs/idea/mcp-tools.md.

The server reads the current directory to pick the project session — open it in your repo and it debugs that repo.

Quickstart

# 1. Scaffold the project (run once in your app's root)
npx @developerz.ai/ui-debugger-mcp@latest init

This creates ./tmp/ui-debugger-mcp/, writes a starter .ui-debugger-mcp.json, and prints the .mcp.json snippet to paste.

// 2. Paste into your project's .mcp.json (add your API key)
{
  "mcpServers": {
    "ui-debugger": {
      "command": "npx",
      "args": ["-y", "@developerz.ai/ui-debugger-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-...",
        "OPENAI_BASE_URL": "https://openrouter.ai/api/v1"
      }
    }
  }
}

// 3. Edit .ui-debugger-mcp.json — set your app's URL
{
  "targets": {
    "web": { "adapter": "browser", "url": "http://localhost:3000" }
  }
}

// 4. In Claude Code (or any MCP client):
start_debug { target: "web", goal: "log in and add item 3 to the cart", url: "http://localhost:3000" }

// 5. Poll until done:
get_findings { session_id: "...", wait: 30000 }

// 6. Read bugs[] + visual[] + summary. Fix code, repeat.

Using it

It's a conversation, not a remote control — five fat tools, not one-per-click:

A run is always time-capped: start_debug's timeout (seconds) overrides the default 300s, so a session can never hang forever — it auto-ends and frees the profile lock when the cap fires.

Logging in — named personas

Almost every interesting screen is behind a login, and re-explaining the login flow in every goal string costs driver steps before the real goal starts. Say it once, in the target:

"targets": {
  "dashboard": {
    "adapter": "browser",
    "url": "http://localhost:5173",
    "auth": {
      "admin": { "path": "/login", "fields": { "email": "admin@dev.local", "password": "admin" }, "submit": "Sign in" },
      "user":  { "path": "/login", "fields": { "email": "user@dev.local",  "password": "user"  }, "submit": "Sign in" }
    }
  }
}

start_debug { target: "dashboard", as: "admin", goal: "open Audit and check the table renders" }

The run signs in out-of-band, before the driver's first step: it costs zero driver steps, and the credentials never enter the model's context — the prompt only learns "you are already signed in as admin". Values are redacted out of every log the run writes; describe reports persona names only. A typo'd as fails loud listing the valid names, and a login that does not take fails the run rather than handing back a session that will report every page behind it as empty. Use dev/seed credentials — this file is committed. Details in docs/idea/config.md.

Telling it what's expected — notes

A freshly-migrated app with no seed data renders empty states everywhere, and the driver reports "the table is empty" as the run's headline bug. It is right, and it is useless. Say what's expected once, in the target:

"targets": {
  "dashboard": {
    "adapter": "browser",
    "url": "http://localhost:5173",
    "notes": "needs seeded data — empty tables are expected on /new\nfirst load shows an onboarding modal; dismiss it\ndark mode is the default theme"
  }
}

One fact per line, on any target (web, desktop, android). Unlike the per-run goal, notes is config: it is composed into the driver's prompt once, as its own "Known about this app" section — treat these as expected, never report one as a defect, and a screen that contradicts one is still worth reporting. Capped at 1000 characters, enforced when the config loads (the prompt carrying it is resent to the model on every step, so an essay here is paid for per step).

If a run is already open

One run per project (cwd), so a forgotten end_session used to wedge it. It doesn't any more: describe reports the run this project holds (session: { id, status, goal }), so a caller that lost its session_id — a compacted context, a restarted client — can read it, close it, or take it over with start_debug({ …, replace: true }). Being refused is still the default, and the refusal spells out all three calls; a silent takeover would kill a healthy run somebody else is watching.

Every tool result carries both a pretty-printed text block and a typed structuredContent payload validated against a declared outputSchema — parse the structured half, don't scrape the text. Tools also declare MCP annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) so clients can render/gate them correctly, and evidence paths (screenshots, replay.mp4, logs) ride as resource_link content items, not inline strings. Full shapes in docs/reference.md.

Typical loop from a smart agent:

start_debug { target: "web", goal: "log in and add item 3 to the cart" }
→ poll get_findings (wait) until status is passed | failed
→ read bugs[] + visual[] + summary, fix the code, start_debug again

You can also drive it headless from a script with claude -p — see docs/claude/SKILL.md for the CLI recipe (MCP config, allowed tools, output formats).

What a run captures (web)

You get findings, but the agent's evidence is what makes them actionable:

Static assets are held back from a default network read — on a Vite dev server they outnumber real API calls ~15:1 — with a count of what was hidden and the filter to see them. Failed requests are never hidden, whatever their type.

CLI — check or stop a run

The ui-debugger-mcp binary doubles as a control CLI for the active run (reads state.json, no API key needed):

ui-debugger-mcp status   # which run is active, server pid, verdict, finding counts
ui-debugger-mcp stop     # gracefully end the run (frees the browser + profile)

Troubleshooting

Chrome not found The web adapter launches Chrome via the system PATH. Install Chrome/Chromium, or set executablePath in .ui-debugger-mcp.json:

"web": { "adapter": "browser", "url": "...", "executablePath": "/usr/bin/chromium-browser" }

Session locked — "another run is active" One Chrome profile = one run. If a previous run crashed without cleaning up:

npx @developerz.ai/ui-debugger-mcp@latest stop   # graceful teardown

Or delete ./tmp/ui-debugger-mcp/<project>/state.json and restart the MCP server.

Run times out with no findings Default cap is 300 s. Raise it per-call:

start_debug { target: "web", goal: "...", timeout: 600 }

If the agent is stuck at login, add ?debug-ai=true to your app's login route (gated by ALLOW_AI_DEBUG_LOGIN) to skip captchas — see CLAUDE.md for the pattern.

get_findings returns empty bugs[] / visual[] The run may still be in progress — use wait (ms) to long-poll:

get_findings { session_id: "...", wait: 30000 }

Check ./tmp/ui-debugger-mcp/<project>/sessions/<id>/logs/agent.log for the agent's trace. Session ids are local timestamps (2026-07-27_14-30-05-0001), so the newest run sorts last. Only the 5 most recent sessions are kept — each new run prunes the rest, so evidence from a run six ago is gone by design.

Run fails instantly: "… is not a valid model ID" The model string in .ui-debugger-mcp.json is not a catalog id. OpenRouter takes provider/model with optional :floor / :nitro routing suffixes — a #… suffix is rejected outright. Use plain ids (deepseek/deepseek-v4-flash).

Still running an old version after upgrading npx -y <pkg> REUSES a cached install when one exists, so a bare "@developerz.ai/ui-debugger-mcp" in .mcp.json keeps booting whatever version first landed in ~/.npm/_npx — with no signal that anything is stale. There is no auto-update: nothing in the server checks the registry. Pin @latest in your args (what init now prints) so every cold start re-resolves:

"args": ["-y", "@developerz.ai/ui-debugger-mcp@latest"]

Already stuck? rm -rf ~/.npm/_npx, then reconnect the server (/mcp in Claude Code). Check what you're on with npx @developerz.ai/ui-debugger-mcp@latest --version.

Config edits seem to have no effect Config is read ONCE when the MCP server starts, so an edit cannot reach a running server. Since v1.3 the server refuses to start a run on a changed file and tells you to reconnect it (in Claude Code: /mcp → reconnect ui-debugger) rather than silently running the old settings.

"Another server already has a session" and you can't see it That run belongs to a different MCP server process on the same project. You can now READ it — get_findings { session_id: "<the id in the error>" } serves its on-disk findings — so you can tell a live run from one that settled and is just waiting to be closed. To take the project over: ui-debugger-mcp stop.

Run ends failed with no bugs and no explanation That means the driver never called report — it used its whole step budget. The summary says so explicitly; it is NOT a crash and NOT a clean pass. Narrow the goal, or raise timeout. Read logs/agent.log and the screenshots for what it actually did.

replay.mp4 not generated ffmpeg is optional. Install it and retry, or ignore — findings and screenshots still land without it.

Stack

• Bun + TypeScript (ships as npm, runs via npx/bunx)

• Vercel AI SDK — the agent loop (fast driver + vision describer)

• Any OpenAI-compatible router (OpenRouter default) — swap models per role. Defaults: deepseek (text) drives, glm (image) sees.

• CDP for web, X11/Wayland for desktop, ADB for Android

• stdio MCP transport

Status

All three adapters ship in v1:

Replay video (replay.mp4, captioned stills → mp4 via ffmpeg) ships with the web adapter. ffmpeg is optional — absent gracefully, findings still land.

See docs/idea/ for design notes.

Docs

• docs/idea/overview.md — problem + idea

• docs/idea/architecture.md — system design

• docs/idea/adapters.md — adapter contract + targets

• docs/idea/desktop-control.md — Linux control tooling (X11/Wayland/mobile)

• docs/idea/agent-loop.md — the story → findings loop

• docs/idea/mcp-tools.md — two tool layers, SQL-like params, in-repo prompts

• docs/idea/models.md — the three actors (smart agent / fast guy / vision guy)

• docs/idea/config.md — config files

• docs/idea/workspace.md — per-project space + logs

• docs/claude/SKILL.md — driving claude as a headless CLI tool (generic)

• CLAUDE.md — instructions for AI agents working on this repo

Credits / influences

• ai-task-master — build template (orchestrator + subagents), reference repo, not published

• gold-standards-in-ai — MCP & code conventions, reference repo, not published

• claude-code-bible — agent-first patterns (sebyx07/claude-code-bible)

• Model Context Protocol