tabmancer

Many Claude Code sessions in many terminals: one screen for all of them, and none lost. tabmancer is a dashboard for your coding-agent sessions, in the terminal. See every session from every project, which ones are working, which one waits for you and how full each one's context is, and go to any of them in a key. Close them when you have to, even halfway through a task: they come back where they left off, each in its own tab and folder, and the ones cut off mid-reply carry on by themselves. Claude Code first, Codex and Antigravity if you use them.

I run Claude Code in several terminals at once, and keeping track of them got hard: which one is working, which one is waiting for me. And I often had to close them before Claude had finished. tabmancer is what I built for both: one screen to see and manage them all, and every session back where it left off, all of them at once. Sinan Tüfekci

Try it

uvx tabmancer --demo         # made-up sessions: nothing of yours is read or changed

Every key works in the demo, but nothing opens and no model is asked. Then install it:

uv tool install tabmancer    # or: pipx install tabmancer
tabmancer                    # the dashboard

It needs Python 3.10+ and Claude Code's claude on your PATH. Upgrade it with uv tool upgrade tabmancer (or pipx upgrade tabmancer): the dashboard says when a newer version is out (updating). ? lists every key and runs the one you pick; the footer shows the common ones and what they do on the row you are on. On Windows, tabmancer shortcut adds a desktop shortcut and tabmancer startup on opens the picker at login when sessions were left open.

What it does

• Shows what each one is doing. Every open session, whichever terminal it is in: working, waiting for you or needing you, the model and effort it runs on, and how full its context is; your plan limits in the top bar. j goes to the next one waiting for you, and Enter on an open one goes to its tab.
• Brings your sessions back. Closed before they were done, by you, a reboot or a crash, they say how they ended (cut off, unanswered, at a usage limit, done) and wait under TO CONTINUE. Enter opens them again where they left off, each in its own tab and folder, and the ones cut off mid-task are told to carry on. After a reboot the ones that were open start out ticked: one Enter brings them all back. Resume, fork or start fresh, per session.
• Gets past a usage limit. Sessions stopped at a limit say when it lifts; c continues them all then, or right away once you have signed in with another account.
• Runs teams. Several sessions on one project with an orchestrator, or as peers on related topics side by side, talking through Claude Code's own messaging between sessions: one row in the list, and a view of who tells whom what. Each member can run on its own model and effort.
• Hands crowded sessions over. h moves the work to a fresh session, with a full handover as its first message. A whole team too: its lead writes the handover and the line-up for the rest of the work, you change what you like, and the team starts again, fresh.
• Goes easy on your plan's limit. Every reply reads the whole conversation again, so a long session is where most of a limit goes. The sessions tabmancer starts keep to a context cap (300k unless you pick another), and in n a quick router suggests the model, effort and cap for your first message: the cheapest that will still do the work well, yours to take with Ctrl+R or leave. The context cap.
• Counts. Tokens, when you work, tools and files, plan limits with a forecast, and a written report of the week to share.
• Wraps your week up. Claude Code Wrapped: your week as a card to post on X, Bluesky, Threads or your Instagram story.
• Finds. / finds a session by anything you or Claude said in it.
• Codex and Antigravity too, if you use them.

Everything it does, and the details: using it (what each status means, every key, writing to a session, background sessions, workspaces) · teams · handing over · stats and reports · setup and commands · how it works

Claude Code Wrapped

Your week with Claude Code as a card worth posting, drawn in pixels like Bit. On Fridays Bit says it is ready, and W shows it any day: who the week made you (the Conductor, the Night Owl, the Marathoner, the Comeback Kid and more, each with the number behind it), the tokens against last week, the time at work, the sessions side by side, what Claude wrote in novels' worth, the files edited, the models, the longest session and the hours you worked. f turns it into a story card. p shares it: X, Bluesky or Threads open a new post with the words already written and the card on your clipboard to paste in; for Instagram the story card is saved and its folder opened, to send to your phone. Nothing is posted until you post it, and the card holds counts only: no project names, nothing said in a session.

tabmancer wrapped --share x    # or bluesky, threads, instagram; --story saves the tall card

Keys

All the keys.

Where it runs

Built and used every day on Windows 11 with Windows Terminal: tabs, going to a session's tab, start at login, desktop notifications and a desktop shortcut. On Linux, tabs open in tmux (tried for real), WezTerm, kitty, GNOME Terminal and Konsole; on macOS in iTerm2 and Terminal.app, which have not been tried on a real Mac yet. In any other terminal, tabmancer prints the commands to run. What has been tried where. Reports are welcome.

Claude Code's formats are its own and can change between releases: after an update, tabmancer doctor checks that everything tabmancer relies on still works.

Privacy

tabmancer runs on your machine and talks to no server of its own. It reads the transcripts the agents already keep (~/.claude, ~/.codex, ~/.gemini) and asks the claude CLI which sessions are open and how full your plan limits are (/usage, a local command). It reads which account you are signed in with (its email and plan) from ~/.claude.json and Codex's auth.json, to show it in the top bar, and remembers the emails of the Claude accounts it has seen, in its own folder, to offer them when you switch (A), with the browser profile each one's sign-in page opens in: for that it reads your browsers' list of profiles (each one's name and the Google account it syncs with), never their cookies or passwords; and the ids of the sessions it opened again, to count them on your week's card (W). Once a day it asks PyPI which version of tabmancer is the latest, to tell you of a newer one (tabmancer updates off stops it). A model is called only when you ask for one, through your own CLI and plan: an AI summary (s), a handover (h; a team's lead writes its team's, h in the team view), a report (r in the stats), a team's line-up (Ctrl+R in the new-team form, which sends the goal, the project's file names and its README's start, and the goal and members' parts of a team at work in that project). One call is made without asking: in n, once you stop typing a first message, the router sends that message and the ticked folders' names to Haiku to suggest a model, effort and context cap ("router": false in its config.json stops it; Ctrl+R then asks). It types into an open session only when you have it continue the ones stopped at a usage limit (c), or send a team at work's lead the message Ctrl+R wrote for it (Ctrl+S). Sharing your week's card (p on it) opens the platform's own page for a new post in your browser, with the words written in: you post it, or not. Nothing is sent anywhere else, and there is no telemetry.

Trademarks

tabmancer is an independent project, not made by, affiliated with or endorsed by Anthropic, OpenAI or Google. Claude Code, Codex and Antigravity are their owners' trademarks; they are named here only to say what tabmancer works with.

License

The code is licensed under the Apache License 2.0: keep the NOTICE with it and credit "tabmancer by Sinan Tüfekci". Bit, the robot, is licensed under CC BY 4.0: wherever you use it, credit "Bit by Sinan Tüfekci". Not affiliated with Anthropic.