MCP
acdev mcp is an MCP server over stdio. It gives AI agents the debugging tools and more: an agent can start an acdev dev session itself, look at the screen, click buttons, read the console, and take screenshots, so it can build and verify your app.
Agents can do all of this with the plain debugging commands too, which need no setup; the acdev skill teaches them how. Use MCP when your agent works better with tools than with a shell.
Claude Code:
claude mcp add acdev -- acdev mcpOther clients: the command is acdev mcp, with no arguments (add --session <pid> to pin a particular session).
| Tool | What it does |
|---|---|
doctor |
Checks this machine for acdev dev, as acdev doctor does; call it first when start fails |
doctor_fix |
Checks, fixes what acdev may fix without asking, and checks again, as acdev doctor --fix does. The fixes that download, install or stop something it only names: the user runs acdev doctor --fix in a terminal (it asks), or agrees to acdev doctor --fix --yes. Its dir must be inside the working directory |
sessions |
Lists running acdev dev sessions; can pin one |
start / stop / logs |
Starts its own acdev dev session (for the project in the working directory by default), stops it, reads its output. When an acdev dev already runs exactly that (the same projects, desktop, app and headless), start uses that session and says so (attached to acdev dev <pid>, which already runs …); stop and logs do not reach a session started elsewhere. When one runs those projects some other way, start fails and names it (… which is not what start asked for): pin it with sessions {use: <pid>}, or stop it first. start {force: true} starts another one anyway |
webs |
Lists pages, and the camera or microphone prompts that wait |
permissions / answer / revoke |
Lists the camera and microphone prompts that wait and the decisions made, allows or denies a waiting prompt, takes back a decision; see Camera and microphone |
snapshot |
The screen as text: windows and the elements of the focused window, each with a ref |
act / run |
Performs one action on a ref / several in sequence, and reports what changed |
open |
Opens an app and waits for its window |
query |
Finds elements by CSS selector, hidden ones too; the refs work with act |
wait_for |
Waits for some text to appear (or disappear) |
screenshot |
Screenshot of one window or the whole screen; the desktop's panels (Dock, menu bar) only on the whole screen |
eval |
Runs JS in a page |
console / network |
The pages' logs and requests |
emulate |
Switches light/dark mode |
choose_folder / choose_files / choose_save_file |
Answers the page's next file picker in advance |
route / unroute |
Fakes responses to the page's requests, or makes them fail |
pointer |
Moves the pointer to a screen point, as a real move (hover, screen edges) |
clipboard |
Reads and writes the clipboard |
input |
Sends raw mouse and keyboard events to a page, at page coordinates; wait (ms) holds a key or button in between |
clock |
Takes over a page's time: installs a fake clock, pauses it, moves it forward |
dialog |
Answers the page's next alert / confirm / prompt in advance, and lists the dialogs it showed |
resize |
Resizes a window so that its page has the given size |
Which session is used: the pinned one; otherwise the one start launched; otherwise one picked by the working directory.
Declared capabilities need no answer; only the camera, the microphone and idle detection (idleDetection) ask the user. The user is asked the first time a page asks for one, and the page waits for the answer. webs lists such a prompt as a line, recorder waits for approval of microphone: the answer tool allows or denies it. answer takes allow (true or false). To allow, it must name the app and the capability that permissions lists, so the call states what is granted; to deny, the prompt id or the app and capability will do, and neither is needed when only one prompt waits. It stands in for the user's own answer, so the tool is marked as an action that needs confirmation: allow only what the user asked for or agreed to.
Closing the only window of a session that runs a single app (act with close on w1) ends that session, by design. The call succeeds and its answer says so, closed w1; acdev runtime exited with it, so acdev dev 4242 ended. The next call fails and says that the session stopped (the acdev dev started here (pid 4242) stopped; last output: … for a session start ran, acdev dev 4242 stopped; no acdev dev is running (…) for one started elsewhere). Run start again to go on. Any other action during which the runtime goes away fails with acdev runtime closed the connection; so does a run whose steps close that window, since the answer to its steps is lost, so close it with act.
input, clock, dialog and resize do for an agent what app.mouse / app.keyboard, app.clock, app.answerDialog and app.resize do in a test; see Time, dialogs and page setup. clock takes action (install, pause_at, fast_forward, run_for, resume, set_fixed_time, set_system_time, state) with a time (an ISO string or milliseconds since 1970) or a duration (milliseconds, "mm:ss" or "hh:mm:ss"). dialog without accept or text only lists.
A page's folder, file or save picker can be answered before the page opens it: call choose_folder with a path (relative paths resolve against the working directory), or with cancel: true, then the act that opens the picker. Each answer serves the next picker of that kind. In a session started with start {headless: true} there is no picker dialog, so a picker that was not answered first fails and the page sees it as cancelled. choose_files takes paths, choose_save_file a path.
{ "name": "choose_folder", "arguments": { "path": "fixtures/project" } }{ "name": "act", "arguments": { "ref": "w1.4", "action": "click" } }On a machine without a display, call start with headless: true; screenshot and snapshot work the same. eval with getComputedStyle or getBoundingClientRect checks styles and layout, as in Styles and layout.
In a session with windows, clipboard touches your system clipboard, so the agent must pass system: true explicitly, and the tool is marked as an action that needs confirmation. Headless sessions have their own clipboard.
When the server's stdin closes or it receives a signal, it stops the acdev dev session that start launched.