A command-line instrument for browser tab orchestration — list, search, group, archive, close, undo — wired into Edge or Chrome through a native messaging bridge. Built for humans who hoard tabs and the AI agents who clean up after them.

Install

npm install -g tabctl
tabctl setup --browser chrome
# Load the extension: chrome://extensions → Developer mode → Load unpacked → paste: ~/.local/state/tabctl/extension/
tabctl ping

If it pings back, the wire is live. You're connected.

Agent Skill

Give your coding agent eyes into the browser. One command and it learns the protocol.

tabctl skill
# or: npx skills add https://github.com/ekroon/tabctl --skill tabctl -a opencode -a github-copilot -a claude-code

Safety

Nothing leaves your machine. No cloud. No telemetry. Just a socket between your terminal and your browser, quiet as rain on neon.

Every mutation is undoable — tabctl undo rewinds closes, archives, and group changes like they never happened. A configurable policy layer shields pinned tabs and protected domains from accidental destruction. You pull the trigger; tabctl keeps the safety on until you mean it.

What You Can Say

When tabctl is installed as a skill, your agent sees what you see. Just talk to it.

"Which of my tabs can I close?" The agent scans for duplicates, stale pages, and tabs you haven't touched in days — then offers to clean house.

"Are any of my open tabs relevant to my note on Project Helios?" When connected to Obsidian, your agent cross-references every open tab against your notes and surfaces the ones that matter.

"I just finished researching service mesh architectures. Organize what I found." Groups your tabs by theme, extracts key URLs, and drops a summary into your notes — before you forget what you were looking at.

"Where's that AWS pricing page I had open somewhere?" The agent searches your open tabs and groups by title and URL — and brings it back into focus.

"Pull every error message from my open Sentry tabs into a markdown table." The agent reads each tab, extracts what you need, and formats it — no copy-paste, no context switching.

"Group everything by project. You know which ones." Your agent infers context from URLs, titles, and your workspace — then sorts ninety tabs into five groups with names that actually make sense.

tabctl works through a lightweight local stack: the CLI talks to a native messaging host, which proxies requests to a browser extension. The host only runs while the browser is open and the extension is connected.

This repo contains:

Chrome/Edge extension (tab/group inspection + actions)
Native messaging host (Node)
CLI (tabctl) for on-demand workflows

Quick Start

1. Build and install

npm install
npm run build
npm link          # puts tabctl on your PATH

If you haven't run npm link, you can always use node ./cli/tabctl.js instead of tabctl.

2. Set up your browser

Run setup — it syncs the extension, tells you where to load it, and auto-derives the extension ID:

tabctl setup --browser chrome

This will:

Copy the extension to a stable location (~/.local/state/tabctl/extension/)
Print the path (and copy it to your clipboard)
Ask you to load it as an unpacked extension in chrome://extensions
Auto-derive the extension ID from the installed path

Edge? Use --browser edge and load from edge://extensions instead.

If you need to override the auto-derived ID (e.g. for a custom extension path):

tabctl setup --browser chrome --extension-id <your-extension-id>

3. Verify and explore

tabctl ping       # check the connection
tabctl list       # see your open tabs

Multiple browsers? See Multi-Browser Setup for running tabctl with both Chrome and Edge.

Commands

Command	Description
`tabctl list`	List open tabs and groups
`tabctl open --url <url> --group <name>`	Open tabs into a group (reuses existing, skips duplicates)
`tabctl analyze`	Find stale or duplicate tabs
`tabctl inspect --tab <id>`	Extract page metadata or CSS selectors
`tabctl group-gather`	Merge duplicate groups with the same name
`tabctl close --tab <id>`	Close tabs with full undo support
`tabctl report`	Generate reports in JSON, Markdown, or CSV
`tabctl undo`	Revert the last action

See CLI.md for the full command reference, options, and examples.

Screenshot output

When --out is omitted, screenshots are written to ./.tabctl/screenshots/<timestamp> and the JSON response includes writtenTo.

Agent workflow (context -> selector)

Use screenshots only when you need visual context, then extract selectors with inspect.

Capture context (full page tiles):

tabctl screenshot --tab <id> --mode full

Identify the element visually, then extract its selector:

tabctl inspect --tab <id> --signal selector --selector '{"name":"target","selector":".your-selector"}'

If you need an absolute URL, set --selector-attr href-url or set attr to href-url/src-url:

tabctl inspect --tab <id> --signal selector --selector '{"name":"link","selector":"a[href]","attr":"href-url"}'
tabctl inspect --tab <id> --signal selector --selector "link=a[href]" --selector-attr href-url

Agent skills

Install the tabctl skill for agents (OpenCode, Claude Code, Codex, etc.) via the bundled command (uses the Skills CLI under the hood):

tabctl skill

This writes a project-local skill to .opencode/skills/tabctl/SKILL.md. You can also install globally:

tabctl skill --global

To install into a specific agent toolchain with skills:

npx skills add https://github.com/ekroon/tabctl --skill tabctl -a opencode

Policy (protect tabs)

By default the CLI loads a policy file from: <configDir>/policy.json (default: ~/.config/tabctl/policy.json)

Set TABCTL_CONFIG_DIR to override the config directory.

This is a protection-only policy that marks tabs as ineligible for agent actions. Example:

{
  "protect": {
    "pinned": true,
    "groupTitles": ["🔒"]
  }
}

Create a default policy file:

tabctl policy --init

tabctl setup does not install a default policy. See config/policy.example.json for a starter template.

Configuration

Config directory: TABCTL_CONFIG_DIR → $XDG_CONFIG_HOME/tabctl → ~/.config/tabctl

An optional config.json in the config directory can set dataDir to override where state files (socket, undo log) are stored. When TABCTL_CONFIG_DIR is set but no dataDir is configured, data defaults to <configDir>/data/; otherwise it uses $XDG_STATE_HOME/tabctl (or ~/.local/state/tabctl).

See CLI.md for full details.

Runtime state

Socket: <dataDir>/tabctl.sock (default: ~/.local/state/tabctl/tabctl.sock)
Undo log: <dataDir>/undo.jsonl (default: ~/.local/state/tabctl/undo.jsonl)
Profile registry: <configDir>/profiles.json

Multi-Browser Setup

Advanced topic — you only need this if you run tabctl with more than one browser (e.g. Edge and Chrome).

tabctl supports multiple browser profiles. Each profile connects to a different browser (Chrome, Edge).

# Setup for Edge
tabctl setup --browser edge

# Setup for Chrome (with custom name)
tabctl setup --browser chrome --name chrome-work

# List profiles
tabctl profile-list

# Switch default
tabctl profile-switch edge

# One-off command with different profile
tabctl list --profile chrome-work

Custom Chrome Profile Directories

If you launch Chrome with --user-data-dir, Chrome looks for native messaging manifests inside that directory. Use --user-data-dir in setup to write the manifest to the right place:

tabctl setup --browser chrome --user-data-dir /path/to/chrome-profile

This writes the manifest to <user-data-dir>/NativeMessagingHosts/ instead of the system-wide location.

How It Works

Each profile gets its own:

Native host manifest and wrapper script
Unix socket for CLI-host communication
Undo history log
Data directory

Policy is shared across all profiles.

Security

The native host is locked to your extension ID.
All data stays local; no external API keys are used.

Development

TypeScript workflow

Source lives in src/ and compiles to build/, then syncs to the runtime locations:

src/extension/background.ts -> extension/background.js
src/host/host.ts -> host/host.js
src/cli/tabctl.ts -> cli/tabctl.js
src/tests/unit/*.ts -> tests/unit/*.js

Build and test:

npm install
npm run build
npm test

Versioning

The base version lives in package.json and is embedded into the CLI, host, and extension at build time.

Commands:

npm run bump:patch
npm run bump:minor
npm run bump:major

Local builds default to a dev version when a .git directory is present, appending the short SHA.

npm run build

This produces versions like 0.1.0-dev.abc12345 (and appends .dirty when the repo has uncommitted changes).

For release builds without SHA, set:

TABCTL_VERSION_MODE=release npm run build

Notes:

close --apply uses the most recent analysis by analysisId.
close without --apply requires --confirm to prevent accidental closure.
Reports include short descriptions from page metadata and a fallback snippet.
list and group-list paginate by default (limit 100); use --limit, --offset, or --no-page.
Use --group-id -1 or --ungrouped to target ungrouped tabs.
--selector implies --signal selector.
Unknown inspect signals are rejected (valid: page-meta, selector).
Selector attr supports href-url/src-url to return absolute http(s) URLs.
screenshot --out writes per-tab folders into the target directory.
tabctl undo accepts a positional txid, --txid, or --latest.
tabctl history --json returns a JSON array in data.
--format is only supported by report (use --json elsewhere).