Agentic Tab Context

Let your coding assistant use the browser tab you already have open.

Keep your website open and ask Claude Code, Codex or Antigravity for help. Agentic Tab Context connects your assistant to that page, with your existing logins. You write requests in your usual assistant chat.

First time here? Complete the one-time installation, including the steps for your assistant.

Watch the 1080p walkthrough · 30-second Safe Local Agents overview

Shared tabs workspace

The extension now ships a real multi-tab side panel: search all tabs, share/unshare without closing them, set the active target, lock sensitive sites, and see live removal when a tab closes. The daemon, side panel, and CLI share one context registry.

• Side-panel prototype — interaction reference
• UX and reliability spec
• Security and privacy review
• Security status after 0.5.0
• Extension ↔ daemon contract
• Workspace implementation plan

atc bindings lists every open context; atc use <ctx> switches the active one. See the command reference.

Everyday use: share, then ask

1. Open your website in the browser you normally use. Log in yourself if needed.
2. Share the tab from the Agentic Tab Context side panel. Select one or more tabs and choose Share selected. Sensitive sites stay locked until you explicitly confirm.
3. Ask your assistant about the page. It starts from atc bindings, the tabs you shared, and can switch the active target or unshare a tab with you.

The assistant cannot bind an unshared extension tab; it gets TAB_NOT_SHARED and asks you to share. Managed tabs opened with atc open are separate and do not need this step.

Optional: point to one exact tab

If you prefer to choose the page yourself, use the small popup in your browser toolbar:

1. On the page you want help with, open Chrome's Extensions menu and select Agentic Tab Context. If its icon is pinned, click that directly.
2. Check the page shown under This tab, then press Let my assistant use this tab.
3. Wait for “Bound. Your assistant can now use this tab.” Return to your assistant chat and say, “Look at the tab I selected.”

The actual Chrome popup, shown with a local demonstration page. This popup selects a page; your conversation stays in your coding assistant.

One-time installation

You need macOS or Linux, Git, Node.js 22 or newer with npm, and Chrome (or another Chromium browser). Open the browser once before installing so setup can find its profile. Windows installation is unsupported and stops before writing files.

1. Install

git clone https://github.com/gainratio/agentic-tab-context.git
cd agentic-tab-context
./install.sh

This builds the CLI, links atc into npm's global executable folder, installs the assistant skill for all three assistants, and registers the browser bridge. Keep the clone in place: the CLI and browser bridge use files inside it.

2. Add the browser extension

1. Open chrome://extensions.
2. Turn on Developer mode (top right).
3. Click Load unpacked and choose the extension folder inside this project.

3. Check the browser connection

atc doctor

The extension.connected array should list your browser. If not, nextSteps tells you what to do. This checks the browser bridge; verify assistant discovery separately using the steps below.

4. Connect your assistant and ask

Follow the activation steps for Claude Code, Codex, or Antigravity below. The installer adds the skill files; your assistant also needs to discover them and be able to run atc. Once connected, ask:

"Look at my browser. The checkout page shows an error after I click Pay. What's wrong?"

The assistant starts from the tabs you shared. The toolbar popup is an optional way to share one particular tab; open it from the toolbar rather than opening popup.html as a file.

Connect your assistant

There are two pieces: the browser extension gives access to your open tabs, and an assistant skill teaches your coding assistant to run atc. The standard installation uses local skill folders; it requires no MCP server or marketplace registration. Run the assistant on the same machine as the browser.

./install.sh already installs all three skill copies. If you installed the CLI earlier and need to add or refresh them, run:

atc setup

Setup prints installedSkills, nativeHostManifests, and extension.folder. It replaces the bundled skill copies at these locations:

Assistant	Skill file installed for your user account
Claude Code	~/.claude/skills/agentic-tab-context/SKILL.md
Codex	~/.agents/skills/agentic-tab-context/SKILL.md
Antigravity	~/.gemini/config/skills/agentic-tab-context/SKILL.md

Claude Code

1. Start a new local Claude Code session after installation.

2. Type /agentic-tab-context to select the installed skill, then ask:

   /agentic-tab-context Look at my browser and describe the page I have open.
   

3. Confirm it runs atc and reports the intended page. If the command is missing, restart Claude Code and check the skill file listed above.

The clone also includes a Claude Code plugin manifest. For an optional, session-only plugin load, launch Claude Code with the clone's absolute path:

claude --plugin-dir /absolute/path/to/agentic-tab-context

That exposes /agentic-tab-context:agentic-tab-context for that session. It still needs the CLI and browser setup above. The standard skill installation is sufficient; the local plugin route has not been exercised in this project's end-to-end tests. See Claude Code's skills and local plugin loading documentation.

Codex

1. Start a new local Codex chat after installation. If the skill is missing, restart Codex.

2. In Codex CLI or the IDE extension, use /skills or type $ and select agentic-tab-context. You can invoke it directly with:

   $agentic-tab-context Look at my browser and describe the page I have open.
   

3. Confirm Codex runs atc and reports the intended page. The desktop app can also use the skill when you explicitly ask it to use agentic-tab-context.

Codex discovers user skills from ~/.agents/skills; no Claude plugin command is needed. See the official Codex skills documentation.

Antigravity

1. Start a new local Antigravity conversation after installation; restart the app if the skill is missing.

2. In the standalone IDE, check Customizations in the agent side panel for agentic-tab-context. In the Antigravity App, invoke /agentic-tab-context, or ask in any surface:

   Use the agentic-tab-context skill to describe the page I have open in my browser.
   

3. Confirm the agent runs atc and reports the intended page.

The installer uses Antigravity's documented global skill folder, ~/.gemini/config/skills. Actual Antigravity discovery and invocation have not been tested in this project. See the official Antigravity skills documentation.

Verify the whole connection

Ask your assistant to run atc doctor, list tabs with atc tabs, choose the intended page by title or address with atc bind, and read it with atc inspect. Success means it reports the actual page's title and visible content. These are commands for the assistant to run; manually choosing a tab in the popup is optional.

If the assistant reports atc: command not found, check command -v atc in its terminal environment. Ensure npm's global executable folder (npm prefix -g, followed by /bin on macOS/Linux) is on that environment's PATH, then restart the assistant. From the clone, node bin/atc.js doctor checks the CLI without the global link.

Things you can ask

• "Look at my browser and tell me why the Save button does nothing."
• "On the localhost tab, fill in the signup form with test data and submit it."
• "Check the console on my open dashboard tab for errors."
• "Take a screenshot of the pricing page I have open."

Which browsers work

Browser	Uses your open tabs?	Status
Chrome, Chromium	Yes, with the extension	Chromium tested automatically on every change; installed Google Chrome 155 tested locally on macOS for toolbar binding, page actions, screenshots, logs, redaction, and close/rebind
Brave, Edge	Yes, with the extension	Same engine as Chrome; not tested yet
Firefox	Extension is written to support it	Not tested yet
Safari (Mac)	Discovery only: can list tabs; binding and actions are unsupported because Safari exposes no stable tab identity	Discovery tested with simulated responses; physical Safari unverified
Chrome, Brave, Edge on a Mac without the extension	Can bind by stable browser tab id; page access needs the Apple Events setting below. No screenshots.	Tested with simulated responses only

In Chrome, Brave and Edge on macOS, the setting is View > Developer > Allow JavaScript from Apple Events. You only need it if you don't install the extension. Enabling Safari automation does not enable Safari binding in atc.

Need a separate, clean browser window instead? atc open chromium (or firefox, webkit) opens one that remembers its own logins.

Is it safe?

• The bridge is local. atc opens no network listening port. The extension talks to atc through native messaging. Your assistant's handling of returned content follows its own settings and terms; normal website requests still occur.
• Password controls are redacted from inspection. Log in yourself, as usual. Values supplied through atc type are passed to the page and removed from error messages. Ordinary page text, console messages and network addresses can still contain sensitive information.
• Your assistant can read whatever is on the tab you give it. Only let it use tabs you're happy to share. Web pages can contain text aimed at AI assistants, so treat what a page says as information, not instructions.

Commands

You rarely need these yourself; your assistant runs them. Every command prints one line of JSON.

Command	What it does
atc tabs	List the tabs open in your browsers
atc bind "localhost"	Connect to the open tab matching "localhost" (also --active, --tab <id>)
atc inspect	Describe what's visible on the tab: headings, buttons, fields, text
atc click --role button --name "Save"	Click something, matched by its exact name
atc type --label "Email" --value "a@b.co"	Type into a field
atc select --label "Plan" --value "Pro"	Choose a dropdown option
atc press Enter	Press a key on the focused element (Enter, Escape, Tab, Space, arrows, Home/End, PageUp/Down, Backspace, Delete)
atc click --role button --name "Delete" --within "row text"	Scope the search to a container that contains that text
atc upload --label "Resume" ./resume.pdf	Attach a local file to a file input (8 MiB cap)
atc screenshot	Save a picture of the tab
atc console / atc network	Show console messages / network requests
atc goto <url> / back / reload	Move around
atc doctor	Check that everything is connected
atc setup	Reinstall the assistant skill and browser connection

Element actions require one exact match. Tab binding prefers a unique exact URL/title match, then a unique active tab; otherwise it lists the choices. Use --tab <id> when you need one specific tab.

Live extension screenshots capture the viewport. --full-page is supported for managed windows and explicitly rejected for live tabs.

The full reference (every flag, the output format, error codes) is in docs/SPEC.md.

Optional: Browser MCP and Chrome DevTools

ATC's extension already provides tab discovery, page actions, screenshots, and console/network capture. That is the default connection described above.

For additional debugging, Chrome DevTools for agents offers performance traces and deeper browser diagnostics. Browser MCP is another way to automate an existing browser through its own extension and MCP server. These are optional, separately configured tools; ATC does not install them.

When using another tool, select and verify the intended page in that tool before taking actions. ATC does not currently map its tab IDs to another connector's targets or hand sessions between connectors. Do not assume their current pages are the same. Combined use has not been tested in this project.

---

Under the hood (for developers)

How it works

flowchart LR
  A["Your assistant"] -->|"runs atc commands"| B["atc"]
  B -->|"local socket, only your user"| C["atc background service"]
  C <-->|"browser native messaging"| D["atc extension in your browser"]
  D --> E["The tab you're using"]
  C -.->|"Mac only, without the extension"| F["Safari / Chrome via macOS automation"]

Loading

The first atc command starts a small local daemon (~/.atc/daemon.sock, authenticated with a per-install token in ~/.atc/ipc.token). It remembers which tabs you shared in ~/.atc/state.json so separate CLI invocations keep working on the same tab context. The Manifest V3 browser extension connects to the daemon via Chrome Native Messaging (com.gainratio.atc) and executes DOM inspection, actions, and redacted viewport screenshots inside shared tabs.

Source layout

Directory	Responsibility
bin/ & src/cli/	atc executable, command parser, atc setup, and Unix-socket client
src/contracts/	JSON-line IPC protocol, typed AtcError codes, upload path-safety policy, and ATC_HOME paths
src/daemon/	Background daemon service, multi-tab ContextRegistry, TTL eviction, and atc doctor
src/drivers/	WebExtension live-tab driver, Playwright managed-browser driver, and macOS JXA discovery fallback
src/native/	Native Messaging host wrapper and browser manifest installer
extension/	Manifest V3 Chrome extension (background worker, popup, multi-tab side panel, content script)
skills/	Assistant skill (SKILL.md) installed for Claude Code, Codex, and Antigravity

Runtime configuration

All settings have safe local defaults and require no .env file for normal use. To override ATC_HOME, custom Chromium binary paths, or context TTLs, inspect .env.example (cp .env.example .env). Secrets and IPC tokens are never configured through environment variables.

Troubleshooting

You see	Do this
atc doctor shows no connected extension	Run atc setup, then reload the extension at chrome://extensions
Extension connected, but the assistant cannot find the skill	Follow Connect your assistant; check its installed skill file and start a new session
Assistant reports atc: command not found	Check npm's global executable folder is on the assistant's PATH; see Verify the whole connection
"blocks page access from other apps"	Install the extension, or turn on the Apple Events setting above
AMBIGUOUS_TARGET	Several things matched. Use atc tabs and atc bind --tab <id>, or add --nth 1 for elements
Safari binding is unsupported	Use tab discovery only, or choose a supported Chromium browser
A reload is required after an update	Reload the affected page, then bind it again to replace the old page agent

Update or remove

To update your clone, stop atc, pull the reviewed changes, and rebuild:

atc shutdown
git pull --ff-only
./install.sh

Reload the extension in your browser and reload previously bound pages after updating, then run atc doctor.

To remove it, run atc shutdown and npm unlink --global agentic-tab-context, then remove the browser extension. Remove the nativeHostManifests and installedSkills files listed in your original setup output. The native manifests are named com.gainratio.atc.json in the browser's native-host folder. Keep ~/.atc (or your ATC_HOME) if you want to retain managed-browser logins and saved screenshots; delete it only when you intend to remove those too. The clone can then be removed.

The npm package is private. Local archives include the compiled CLI and extension assets; publishing an npm package or Chrome Web Store build is a separate release action.

Contributing

npm run gate               # build, type-check, and run all tests (including a real browser)
dagger call quality sync   # the same gate inside a pinned container, as CI runs it