Pair with the agent, as a team
A repo, a conversation with the agent, a shared terminal and a set of changes. Everyone sees the same thing as it happens.
See a session
Multiplayer coding agents
Work with the agent together in a shared session, or hand Tandem an idea and let it plan, build and merge the whole thing.
Source code coming soon
A repo, a conversation with the agent, a shared terminal and a set of changes. Everyone sees the same thing as it happens.
See a session
Tandem designs it with you, splits it into tasks, builds them in parallel, checks each one and merges them in order. You stay at the gates.
See how plans work
Every image here is the running app with seeded demo data, in a workspace for Northwind, a made-up web shop with four people on it. Switch the theme with the button above.
Explorer, editor and the conversation in one window. Every edit, cursor and file mark says who made it: a teammate, or Claude.
Switch to the chat-only view when you don't need the code. The history gets the width, and what needs you stays loud.
Teammates join a session to watch the agent work, prompt it, or take over with their own hands. Presence is live, and so is the terminal.
The Changes view lists what changed, shows the diff and holds the commit box. Commit, push the branch and open the pull request without leaving the session.
The agent runs real commands: installs, tests, your scripts. In Tandem they run in a Docker container that belongs to one session, never on the server itself.
Send a prompt as a team. The lead proposes subtasks and the files each one will touch; you edit the list and start them. Each subtask gets its own branch, worktree and tab.
For work bigger than one session. Describe the idea; Tandem designs it with you, breaks it into tasks, builds them in parallel, checks each one and merges the result. You stay at the gates.
It reads the repo and asks the questions that change the design, with a recommended answer for each. Anyone on the plan can answer.
The architect proposes tasks with their dependencies, a model for each and an estimate. Change any row, then approve to start building.
Ready tasks start by themselves, up to the limit you set. Finished ones wait in the merge queue, which rebases, runs your checks and pushes.
Each task has acceptance criteria. A verifier reviews the result against them and reports evidence per criterion. If one is unmet, the task goes back for another round, and the cost is counted against the budget.
An automation runs a prompt on a repo every night, every Monday, or when you press a button. It runs as its owner, in its own session, and says so.