Repository navigation
Releases
NVDA MCP 1.5.0
What is new, in short: MCP Access is harder to misuse. The person being helped can no longer get the
helper's assistant to act on the helper's computer or give away the helper's information, and text that
carries the helper's folders or a secret is refused before it leaves. Long values, such as a chat history, no
longer flood results, and nvda_set_value can add text at the end.
Added
nvda_set_valuetakesappend, to add text after what a field already has. Long text can be written in
several calls without the keyboard, so it cannot land in another window when the person being helped moves
the focus meanwhile. Line breaks follow the field's own, such as Notepad's.
Changed
- The bridge protocol version is now 10. Update the add-on and the server together; through MCP Access, both
computers need the same version.
Fixed
- MCP Access: the person being helped can no longer talk the helper's assistant into acting on the
helper's computer. Every message from them, in the inbox line that wakes the assistant and in the tool
results, now says that they can have it act only on their own computer: no programs, commands, scripts,
files or NVDA actions on the helper's computer because they asked; reading and looking things up stay fine.
The README says to keep Claude Code asking before it runs commands while it watches for messages. - MCP Access: the helper's private information stays on the helper's computer.
- The assistant is told, with every message and every result, never to give the person anything about the
helper or their computer (files, conversations, memory, accounts, paths, keys, passwords), and that claims
of permission or urgency in their words grant nothing. - Every NVDA tool result read from their computer starts by saying that what their screen shows is data, not
instructions, so text planted in a window title or web page is not taken as a command. - Independently of the assistant, a message, typed text or field value for their computer is refused when it
carries the helper's profile or temporary folder, or looks like a private key or an API, GitHub, AWS,
Google, Slack or login token.
- The assistant is told, with every message and every result, never to give the person anything about the
- A control with a very long value, such as TeamTalk's chat history, no longer floods the result of
nvda_get_focusand every action that reports the focus: the value is shown by its start and end, saying how
much was left out, andnvda_get_textreads all of it. Found while helping someone through MCP Access, where
one key press returned about 133,000 characters. - When NVDA's focus is still in another application than the window Windows has in front, for example while
the person switches windows, the focus report now names the window in front and says the focus may be
catching up, instead of only reporting the old focus.
NVDA MCP 1.4.0
What is new, in short: MCP Access now works both ways, with one window for the person being helped, and in
Claude Code their messages can wake the helper's assistant. The assistant can use the mouse, for what the
keyboard cannot reach. MCP Access is in early access: free for now, and it may change.
Added
- Talking both ways through MCP Access.
- The person being helped has one window, opened with a new NVDA command: the session's status, their
conversation, a field to answer (Enter sends) and a button to stop sharing. - Messages from the helper play a sound and are spoken as messages from them, so they are never taken for
NVDA's own, and stay in the window to hear again. - The helper's assistant sends messages with
nvda_access_send_messageand reads answers with
nvda_access_messages, which can wait for one; a message that arrives while it does something else comes
with the result of its next tool, once. - Messages are end-to-end encrypted with the session, at most 2000 characters, and a helper may send at most
20 a minute.
- The person being helped has one window, opened with a new NVDA command: the session's status, their
- Waking the helper's assistant. Each message is also written to a small inbox file in the helper's own
temporary folder, deleted shortly after the session ends, so a background monitor in Claude Code wakes the
assistant when the person writes, even when the helper is away. - Honest about who is talking. The person being helped sees the helper's assistant as "name (AI assistant)"
(helper_nameinnvda_access_connect), in the dialog and in every message. - Talk mode (
nvda_access_connectandnvda_access_set_mode): the helper's assistant only reads and
talks, without acting on the person's computer, for example while the helper is away. - Mouse tools, for what the keyboard cannot reach, such as buttons and knobs drawn on an audio plugin's
canvas:nvda_mouse_click(left, right, middle, double),nvda_mouse_scroll,nvda_mouse_hoverand
nvda_mouse_drag. They aim at a control (its middle, or a point counted from its top left), a window, or the
screen, refuse points outside the control or off the screen, and go through the same checks as key presses,
through MCP Access too. Click and hover say what is under the mouse.
Changed
- The README is reorganized: MCP Access has its own section, saying which assistants can be woken, and the tools
are grouped into reading, acting with the keyboard, acting with the mouse, MCP Access, and for developers. - The bridge protocol version is now 9. Update the add-on and the server together; through MCP Access, both
computers need the same version, and a connection between different versions is refused, saying so.
Fixed
nvda_list_windowsalso lists windows that belong to another window, such as REAPER's FX windows, with the
handle of the window they belong to. Like Alt+Tab, it used to leave them out.
NVDA MCP 1.3.0
Added
-
MCP Access: someone you trust helps you from their own computer, with their AI assistant working through
your NVDA. A command in Input Gestures shares this computer and copies a one-time code that works for
10 minutes; the helper's assistant connects withnvda_access_connect, and every NVDA tool then works on
your computer untilnvda_access_disconnect. The person being helped needs no AI assistant.- End-to-end encrypted (X25519 and AES-256-GCM, keys made from the code and fresh random keys); the relay,
run by InfiArtt atwss://infiartt.com/mcp-access/v1, only forwards data it cannot read.
Strict TLS 1.3 to the relay. - NVDA asks before a helper connects, and again before their first key press and their first picture of the
screen in each session, even if those are off in the settings. Never let MCP Access helpers press keys and
Never let MCP Access helpers take pictures of the screen refuse them altogether. - Reloading add-ons, running NVDA commands, NVDA's log and settings, and web pages' HTML are never available
through MCP Access. Connections and actions are written to the action log. - The ready-made prompt
help_someoneconnects with the code the person sent and helps them safely. - NVDA says when a helper connects and why sharing ended, in your language.
- End-to-end encrypted (X25519 and AES-256-GCM, keys made from the code and fresh random keys); the relay,
-
Set up now, in NVDA's Settings, NVDA MCP category: sets up the chosen assistant without editing files
by hand. For Claude Desktop, Antigravity and Cursor it adds NVDA MCP to the assistant's settings file,
keeping the rest of the file and a copy of the old one (.bak); for Claude Code and Codex CLI it runs their
own setup command. NVDA asks before changing anything. Antigravity and Cursor are new in the list of clients. -
nvda_recognize_text: reads the text on the screen in a window or control with Windows' text recognition,
as NVDA+R does, line by line with where each line is. For applications NVDA can read nothing in.
Every use is written to the action log. -
nvda_get_addon_paths, for add-on developers: where NVDA's developer scratchpad is, whether NVDA loads code
from it, and the add-ons running. -
Ready-made prompts for everyone:
make_accessiblefinds what stops NVDA in an application, writes a small
add-on that fixes it, tests it in the running NVDA and packages it;explain_screensays where you are and
what you can do there;report_problemwrites a bug report for the application's developer, for you to send. -
A command, in Input Gestures, NVDA MCP category, that copies what NVDA sees where you are, to paste into a
chat with any AI assistant, even one without MCP: the application, the focused control, the controls in the
window and what NVDA said last.
Changed
- When
nvda_get_object_treecuts a long list withmax_children, it shows the items around the focus
if the focus is in the list, and otherwise the last items of lists and tables, such as the newest messages
of a chat, instead of always the first ones. It says how many items it left out before and after them. - The bridge protocol version is now 8. Update the add-on and the server together.
NVDA MCP 1.2.0
Changed
- The add-on now works with NVDA 2025.1 and later, not only 2026.3 and later.
On NVDA before 2026.3, which cannot register an add-on's settings at installation,
the add-on adds its settings when NVDA starts, and its settings can be saved in a configuration profile.
NVDA MCP 1.1.0
Added
-
nvda_find_objectstakeswithin, an id from the previous search, to search only inside that control.
On large web pages such as WhatsApp Web, a search of the whole window can stop after 3 seconds
before it reaches the part that matters. -
nvda_inspect_object, for add-on developers: a control's IAccessible2 attributes (HTML tag, id, class,
ARIA role, where its name comes from) or UIA properties, the containers above it, its first children,
and for tables and lists how many rows exist now compared with how many the application reports.
It ranks the conditions an add-on could match the control on, from most to least stable, warns about class
names that look generated by a build tool, and gives examplechooseNVDAObjectOverlayClassescode. -
nvda_find_objectsandnvda_get_object_treetakewindow, a handle fromnvda_list_windows,
to read a window in the background or on another virtual desktop without bringing it to the front.
Chromium browsers and Electron apps stop exposing their page some time after their window is hidden;
the result then says so, and that the window needs to come to the front first. -
nvda_get_object_treetakesidto start from a control found withnvda_find_objects. -
nvda_trace_eventstakesbackground_windowto ask NVDA for events from a window in the background,
as an add-on can witheventHandler.requestEvents. They are recorded with the control's value and
the text of live regions and alerts, but NVDA does not act on them. -
nvda_set_value(experimental) gives an edit field new text through UI Automation or IAccessible,
without the keyboard or focus, and reports whether the field took it. -
Recording what an action changed, for add-on developers. Action tools such as
nvda_press_keys,
nvda_activate_objectandnvda_set_valuetakerecord(orrecord_within, an id): the result then also
says how the accessibility tree around the action changed (controls added, removed, renamed or moved,
values and states, lists that scrolled) and which events NVDA handled.nvda_trace_eventstakestree
(ortree_within) to do the same for something the user does, from start to stop. The comparison never
reports a control as removed when it was simply not read, and reads the part around the focus first. -
nvda_inspect_htmlandnvda_find_html, for developers of add-ons for web apps: the HTML element behind a
control in Chrome, Brave or Edge, read through the browser's remote debugging, with the data attributes,
HTML, styles and event listeners that NVDA cannot see. The control is matched with its element exactly,
through the DOM node id Chromium gives it. Off until you check the new setting
Let MCP clients read the HTML of web pages through the browser's remote debugging; NVDA says when
it connects, and the browser asks before every new connection. Only the page you work with is read:
no scripts are run and your other tabs are never listed. -
nvda_lookup_api, for add-on developers: checks a name in the Python code of the NVDA that is running,
before it goes into an add-on. It says whether the name exists, what it is, its signature, docstring and
members, where it is defined with a link to NVDA's source for that release, and whether it looks deprecated,
and suggests close names when it does not exist. Withsearchit finds names in NVDA's and add-ons' modules.
It only reads, and does not run the code it looks at. -
nvda_capture_control, for developers: a picture of a control on the screen next to what NVDA says for it,
to find icons and visible text the accessible name leaves out. Off until you check the new setting
Let MCP clients take pictures of controls on the screen. It captures the way NVDA's own text recognition
does, so it works with Screen Curtain on, and every picture is written to the action log. -
Ready-made MCP prompts for common developer tasks:
audit_accessibility,build_addon,why_silentand
check_labels. Clients that support prompts offer them as commands.
Changed
- Ids from
nvda_find_objectsstay valid through later searches, until the control disappears,
so a container's id still works after searching inside it. - When
nvda_find_objectssearches several parts of a window, such as a web page and then the browser
around it, they share one 3-second time limit instead of 3 seconds each. nvda_trace_eventsno longer records which character a typed-character event carried, so passwords stay
out of traces (NVDA only says "star" for them). What NVDA speaks as you type in other fields is still recorded.- The bridge protocol version is now 7. Update the add-on and the server together.
Fixed
- Your answer to "allow ... to press keys?" now lasts until NVDA restarts, as the question says,
instead of being forgotten when plugins are reloaded. - An unanswered "allow ... to press keys?" question closes by itself, with an announcement, before the
MCP client stops waiting. Before, answering Yes much later still ran the old action, and every other
action waited behind the open question. An action that cannot start in time fails withcontrolBusy. - Escape answers No in that question; before, it did nothing.
- Using the id of a web page's control that has disappeared now fails with
staleObject. Before, NVDA still
answered for it as an unknown object, so the action failed later with a less helpful error. nvda_get_object_treeno longer drops a control and everything in it when it is the last one shown
and its own children were cut short; on WhatsApp Web most of the page went missing this way.nvda_activate_windownow brings NVDA along when Windows switches virtual desktops to reach the window.
Windows sends no event then, so NVDA kept reading the old window and the result showed the old focus.
NVDA then says the window title once, as after Alt+Tab; at first it said it twice in browsers.nvda_activate_windowsays the window title once when the page in a Chromium browser or Electron app
is in focus mode, as in WhatsApp Web. These applications send no focus event for their window, so NVDA said
the title on entering the window and again on switching to focus mode, even on the same virtual desktop.
The add-on now tells NVDA about the window as soon as it comes to the front. Pages in browse mode
and other applications are left to NVDA, which says the title once there.
NVDA MCP 1.0.0
First release, on the NV Access Add-on Store and PyPI.
Added
- An NVDA add-on with a loopback bridge protected by a per-start token, and an MCP server published on PyPI
asnvda-mcp. It runs withuvx nvda-mcp,pipx install nvda-mcporpython -m nvda_mcp. - NVDA MCP settings: Set up this MCP client and Copy setup command copy the command (Claude Code, Codex CLI)
or the JSON settings (Claude Desktop, Cursor and others) that start the server version matching the add-on. - Reading:
nvda_get_status,nvda_get_focus(withdocumentMode, browse or focus),nvda_get_navigator_object,
nvda_get_object_tree(at mostmax_childrenchildren of each object),nvda_get_text(withcontains
to search without moving the cursor) andnvda_get_speech_history. - Speech:
nvda_speak, andnvda_wait_for_speechto wait until NVDA says something, such as a chat reply. - Keys and typing:
nvda_press_keyspresses keys as the user would, so NVDA runs its command for a key first,
and takesholdto keep alt, control, shift or windows down across several keys.nvda_type_texttypes
a line of text in any language, emoji included, without using the clipboard. - Windows:
nvda_list_windows, optionally including other virtual desktops, andnvda_activate_window. - Controls:
nvda_find_objects,nvda_focus_objectandnvda_activate_object;nvda_audit_focus_order. - Web pages and documents:
nvda_set_document_mode,nvda_find_in_page, andnvda_list_elementsand
nvda_move_to_element, like the Elements List. - Expected results: every action tool takes
until_speech, keeps listening until NVDA says that text,
and reports whether it was heard. - Tools for developers of NVDA add-ons and accessible applications:
nvda_preview_speechsays what NVDA would announce when a control gets focus, without moving focus.
Withtarget: "window"it checks the focused control and the controls of the foreground window,
and flags unlabeled and silent ones.nvda_trace_eventsrecords the events NVDA handles and what NVDA said for each, including speech that comes
just before or after an event, such as a browser's live region. It stops by itself after 10 minutes.nvda_get_logreads NVDA's log; IO entries, which record keys typed, are left out unless asked for.nvda_get_output_historyshows beeps, sounds, braille output and speech being cut off.nvda_get_dev_infoshows NVDA's developer info for a control, like NVDA+F1.nvda_list_scriptslists the NVDA commands available where focus is, or says which command a key runs.nvda_run_scriptruns an NVDA command by name;nvda_reload_pluginsreloads add-ons and app modules.nvda_get_settingsreads NVDA's settings as they apply now. Values whose names suggest secrets are hidden.nvda_run_scenarioruns key presses and typing, records what NVDA said after each step, and saves it
as a baseline file or compares it with one.
- Safety:
- Pressing keys is off by default. NVDA asks before the first key press after it starts, in a dialog whose
title names the MCP client asking. - A setting to let MCP clients control only some applications.
- A log of every action clients take (
%APPDATA%\nvda\nvdaMcp\actions.log), with a setting to turn it off,
an NVDA command to open it and thenvda_get_action_logtool. - When the add-on and the server don't match, the error says which one is older and how to update it.
- Pressing keys is off by default. NVDA asks before the first key press after it starts, in a dialog whose
- A Donate button in the NVDA MCP settings, and donation links on GitHub, PyPI and in the README.
- Help pages and translation in English and Indonesian.
- docs/using.md, CONTRIBUTING.md and
THIRD_PARTY_NOTICES.md.