Gaddi
Provides a browser_signin capability that signs agents into websites by filling credentials from 1Password using its CLI and desktop-app integration. The user authorizes the target site and Login item, Gaddi fills them and returns an outcome to the agent; credentials are excluded from sign-in results and audit entries, and exact matches are redacted from text tool output for ten minutes after use.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Gaddi is a local MCP server that lets AI agents work in your existing Chrome tabs.
Agents use your existing logins, open background task groups, read pages and interact with controls.
I built it to work with MCP-capable agents without taking focus from my other apps.

Read with OmniRead
OmniRead is Gaddi’s reading companion. Gaddi captures the page in your signed-in tab; OmniRead extracts Markdown, indexes its sections and reports evidence of missing content. Together, an agent can read a subscriber article, request one section and distinguish a full capture from an abstract or blocked page. A completeness verdict is evidence, not a guarantee.
Related MCP server: Umbra MCP Server
You authorize the action
A held action waits for you to confirm it in Gaddi's Mac panel. Telling the agent "yes" does not create that signature.
A local broker checks each request before Chrome. The default policy holds recognized payments, purchases, deletion, security changes and sends, plus all uploads. Gmail Send and direct password-field access are refused.
The panel names who asked (the agent, its chat and project), the action, the agent's one-line reason and what approving allows, and can show the target. Agents pass the reason as why; see who asked and why. Touch ID or your Mac password authorizes a Secure Enclave key to sign your decision. Approval binds to the pending action; changed requests and reused proofs fail. The button labelled “Approve with Touch ID” also accepts the Mac password. There is no software-key fallback.
Sign-ins, eligible sends and uploads can remember a site after a signed decision. The shipped lists are empty. Remembered uploads still hold on protected pages and refuse protected files. Payments, deletion and security actions cannot be remembered. See authorization details.
Grant one chat a few holds
An agent can ask once, at the start of a chat, for a session grant: a short list of held actions it may then take without asking, each tied to an address prefix.
browser_grant rules: ["upload https://github.com/settings/profile", "post https://www.linkedin.com/company/123/admin"]
minutes: 180 label: "Update the org profile and post the launch note"Gaddi shows the rules, the length and the agent's label on one card. The grant exists only after you approve that card, with Touch ID, or from your phone if you turned that on (below). It covers the chat that asked and no other, and ends when the time runs out, when that chat closes, when you press End now in the panel, or when the broker restarts. Only uploads, posts and deletes can be granted. Payments, purchases, sends, sign-in and security changes stay held every time, and a button whose name mixes a granted verb with another hold verb stays held. Keep prefixes narrow: a grant on github.com/settings would also cover its key and token pages. Details are in authorization.
Approve from your phone (off by default)
A local bot on your Mac, such as a Telegram bot, can answer the same small set of requests from your phone: uploads, posts and deletes on pages that are not protected, and session grants of up to two hours. You turn it on once in Gaddi's panel with Touch ID, which trusts one Ed25519 key the bot published; anyone can turn it off, from the panel, gaddi remote off or the bot. Each phone answer is signed, bound to one request and usable once, and the panel lists recent ones. Payments, purchases, sign-in, security pages and sending messages stay Touch ID only. The bot's key is a file your user account can read, so another program running as you could approve these requests too; that is why the scope stays small. The contract is in authorization.
Where that protection ends
Gaddi guards its own calls. An agent with shell access can act outside it. Policy matches control names and URLs, mainly in English and Italian; it can miss harmful actions and obfuscated JavaScript. Pages can send requests on their own.
Private information visible in logged-in tabs goes to your agent. The extension needs broad site, debugger, bookmark and extension-management access. Review the default policy and bridge limits.
Install
Requires macOS 26+, a Mac with Secure Enclave, Chrome 120+, Node.js 22.18+ on the 22.x line or 23.6+, and Xcode 26+ to build the approval app. This is a source installation with an ad-hoc-signed Mac app.
git clone https://github.com/ABCastor/gaddi.git
cd gaddi
npm ci
bash install/register/daemon.sh
bash install/register/app.sh
bash install/register/chrome-bridge.shIn chrome://extensions, enable Developer mode and Load unpacked from extension/. Add ~/.local/bin to your PATH. Keep the checkout: installed components reference it. The Chrome bridge needs node on its non-login PATH; version-manager installations may need an accessible binary in /opt/homebrew/bin or /usr/local/bin.
Quit your client before registration, then restart it.
Client | Command |
| |
| |
| |
| |
Antigravity |
|
Other clients can launch node /absolute/path/to/gaddi/mcp/server.mjs over stdio. Installers accept --dry-run. See configuration and removal.
Work in a tab
gaddi status
gaddi open https://example.com --group "Research"
gaddi look <tab-id>
gaddi click <tab-id> '<element-ref>'
gaddi close <tab-id>open returns a tab ID; look returns text and control references. Keep the ID through the task. New tabs use Gaddi's own window, created minimized. You can open it from Chrome's Dock menu; Gaddi leaves its size, position and state as you set them. MCP tools use a browser_ prefix and also support typing, scrolling (sideways too, with dx, and inside a carousel or grid when you name it), selection, waits and screenshots.
A click can go where you point instead of at a reference: gaddi click <tab-id> --at 120,240 (at:{x,y} over MCP) presses that point in the tab as your own mouse would, which is how a button under a cookie banner, or a control inside a third-party embed, is reached. look also lists every document inside the tab with its frame id; --frame <id> (frameId) makes look, click, type, press, hover, select, scroll, wait and upload work inside that document, cross-origin included, and keys go in once that document is the inspected target and holds the focus. The gates follow the click: a point is judged on the control under it, a click inside an embed is judged on that site as well as on the page, a file sent to a field inside another site's embed asks every time (a remembered upload site covers its own pages only), a wheel inside a document needs a selector to aim at, and a point over something Gaddi cannot inspect waits for you. Verification widgets are named as such in look; pressing their checkbox is your own click, and Gaddi never answers a puzzle. Rendering actions run in background tabs. browser_show selects a tab and requests your attention without taking window focus by default.
look returns positions for its outlined references: {ref, x, y, width, height} in the inspected document's viewport, in CSS pixels. Use these to associate controls with visual rows; the outline keeps semantic document order, which may be the reverse of screen order. Inside an iframe these are not the parent tab's click coordinates. Geometry does not establish that an embedded document is on screen or authorize a click. Native select is a convenience that dispatches synthetic events; it does not promise isTrusted input.
Pass the version from look to reject stale input. When a session ends, its background tabs are cleaned up; tabs shown to you are preserved.
browser_read uses OmniRead to extract Markdown, sections and completeness information from page HTML, including logged-in tabs. Install it before registering your client, or set OMNIREAD_BIN. Without it, Gaddi returns page text without sections or continuation handles.
Sign in with 1Password
browser_signin uses the 1Password CLI and desktop-app integration, without needing the browser's password-manager extension. You authorize the site and Login item; Gaddi fills it. A locked or unavailable CLI returns an unlock/retry instruction after a five-second process budget. Agents can work through visual challenges with screenshot/click; passkeys and extra verification can need you.
For a new account, pass create:{title,username,vault}. You approve the new item and vault; op generates and saves its password, then Gaddi fills the signup form without submitting. Use signup:true with item to fill an already saved login. Submit with an ordinary click, keeping its existing approval rules. See the sign-in contract.
Credentials stay out of sign-in results and audit entries. Text output redacts exact matches for ten minutes after use; screenshots do not. Credential-bearing URL query/fragment values are redacted independently, across results, errors, events and stored broker metadata. Navigation and signed approvals still use the original URL. Existing broker state and audit files are scrubbed at startup; client transcripts and Chrome history are not. See URL credentials.
Uploads have a 20 MB limit. Protected credential paths and key files are refused; arbitrary files can still contain secrets. See extension controls. Connection recovery does not resubmit pending page actions.
When Chrome stops
If Chrome has quit or crashed, the next browser call reopens it in the background with your normal profile and last session (open -a, never a fresh profile), waits up to 45 seconds for the extension, and tells the agent and you once (a Gaddi banner) what happened. It reopens Chrome at most once per 10 minutes and three times an hour, for a stopped browser. If three consecutive bridge requests time out within five minutes, or the native host disconnects for at least 15 seconds, browser_status reports stalled. On the next browser call, Gaddi asks Chrome to quit gracefully, waits up to 20 seconds for it to exit, then reopens your normal profile and last session. If Chrome refuses to quit, recovery stops without force-killing it. The extension must reconnect and answer a tabs request before recovery succeeds. Gaddi attempts this restart at most once per 30 minutes, remembers the limit across broker restarts, records the reason in its audit, and shows the outcome in status and a Gaddi banner. Nothing reopens Chrome until an agent needs it.
Development
npm run typecheck
npm run build:extension
bash tests/run-all.shTests use isolated state and falsify gates by removing their protection. Browser tests need Chrome for Testing; Secure Enclave checks need a desktop session. Missing prerequisites are skipped. Commit regenerated extension output after changes. Visual sources are included.
browser_click accepts download:"/absolute/path/export.zip" for a same-origin blob/filesystem anchor export that reaches the top document. It saves a new file under Downloads or the system temporary directory, up to 20 MiB, with no overwrite. The capture does not ask for approval unless hold.downloads is true in your policy file; the click's own holds still apply. Detached anchors, other download methods and embedded documents are unsupported. See the bounded download contract.
When an agent's click or Enter in a tab you are not looking at makes the page open a new tab or popup, Gaddi opens that address itself as a background tab, because Chrome brings its window in front of your work for any tab a page opens. A page that needs its real popup (a sign-in window that reports back to it) gets one when that tab is in front of you.
There is no console or network inspection, general download manager, browser-dialog handling or video recording. CSS and Web Animations can be slowed; JavaScript timers, canvas and video cannot.
Original code is licensed under Apache-2.0. Bundled fonts, the Public Suffix List and dependencies keep their own licenses; see NOTICE.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Let AI agents work in your own signed-in Chrome, with approvals and verified results.
Let AI agents use your real, signed-in Chrome: tabs, clicks, forms, screenshots and video.
Let AI agents use any website you're logged in to. Sessions stay encrypted on your machine.
Let your AI agent operate any website through a real, logged-in browser session.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceLets AI assistants control your real Chrome browser to perform web tasks like reading pages, taking screenshots, clicking, and typing, using your existing logged-in sessions.145MIT
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to securely control a user's existing signed-in Chrome browser through isolated tab groups, with strict per-session ownership and no cookie or token exposure.6 npmISC
- AlicenseNot gradedqualityCmaintenanceConnects AI agents to your existing daily-driver Chrome so they can inspect, extract from, record, and replay workflows across background tabs without stealing cursor focus, breaking 2FA/SSO sessions, or leaking secrets. A deterministic R0–R4 risk firewall gates state-changing and destructive actions behind human approval.2Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to drive the Chrome, Brave, or Edge browser you are already logged into, working in a tab of its own with a visible cursor and real trusted input. You can pause, take over, or hand back control at any time, and consequential actions such as submitting, paying, sending, deleting, or uploading wait for your approval.1Apache 2.0