# How to choose a multiplayer AI coding workspace

Feature lists hide the choice. These nine questions expose it, and the answers sort every tool on the market.

Published 2026-09-16 by the Poly team. Canonical: https://usepoly.co/how-to-choose-a-multiplayer-ai-coding-workspace
Poly (usepoly.co) is a web app where a team builds with AI, Claude or ChatGPT, together in a shared room. Free to start: https://usepoly.co/

**Choosing a multiplayer AI coding workspace comes down to nine questions, and the first one decides most of the rest: does the team share a live agent session, a project with separate agent threads, or an agent in the team chat? After that: who can approve a change before it happens, whether each turn says who spent what, whether a person who cannot code can take part, whether anyone has to hold API keys, where the agent runs, which agents it runs, what the output is, and how it is priced.** Poly (usepoly.co) is one answer to those questions, not the answer to all of them; this page gives you the questions and what a good answer looks like for your team.

*The nine questions and what a good answer looks like.*

|  | What to ask | A good answer for a small team |
| --- | --- | --- |
| Shape | Live session, parallel threads, or agent in the chat? | Live session, if you want a shared record of decisions |
| Approvals | Who can say no, and before or after the change? | Any member, before |
| Cost | Does each turn say who ran it and what it cost? | Yes, per turn, per person |
| Non-coders | Can they prompt, approve and undo from their screen? | Yes, without a terminal or key |
| Keys | Who holds them and can the sandbox see them? | Nobody; never inside the sandbox |
| Runs where | Laptop, your VM, or hosted sandbox? | Isolated per run, allowlisted network |
| Agents | Real Claude Code and Codex, own agent, or many models? | The ones your team already knows |
| Output | A running app, a pull request, or a reply? | The app, with commits behind it |
| Price | Seat, per AI user, credits, hours, or flat? | Free for people who only watch and approve |

## 1. What does the team share?

The three shapes are explained in [what a multiplayer AI coding workspace is](/what-is-a-multiplayer-ai-coding-workspace). Live steering means one session everyone watches and acts on, which gives a small group a shared record of every decision. Parallel threads mean each person's agent works alone and the team shares the merge, which scales to many engineers. An agent in the chat means requests and pull requests, which suits a whole company that already lives in Slack. Decide which one you want before you look at a single feature, because every other answer follows from it.

## 2. Who can approve, and when?

Ask whether a member other than the one who prompted can say yes or no to a change, and whether that happens before the agent makes the change or after a task is finished. Most products answer "the person who started the session" and "after". A pair or a team with non-engineers wants "any member" and "before".

## 3. Does each turn say who spent what?

If the tool runs on each user's own subscription or API key, it structurally cannot tell the room who ran what for how much; the spend is private to each account. If it bills usage itself, ask whether the receipt is per turn and per person or a pooled balance. Attribution matters more the more people share the agent.

## 4. Can someone who cannot code take part?

The honest test is whether that person can prompt, approve and undo from their own screen without a terminal, a git command or an API key. Viewer seats and read-only links do not pass it. Terminal-first tools do not pass it either, whatever their persona pages say.

## 5. Does anyone hold keys?

Bring-your-own-token products require each participant to connect a Claude or ChatGPT plan or paste an API key before doing anything, and keys on a shared machine are readable by every teammate and every instruction in the repository. Ask where the keys live and whether the agent's sandbox can see them.

## 6. Where does the agent run?

On a member's laptop, on a VM the team owns, or in a hosted sandbox. Each is a legitimate choice with a different trust story; the [secure workspace guide](/secure-shared-coding-agent-workspace) walks through it. The questions to ask are whether each run is isolated, whether the network is allowlisted, and whether a new domain needs a human.

## 7. Which agents and models?

Some workspaces run the real Claude Code and Codex agents; some run their own agent; some run dozens of models through an open harness. Decide whether you want the agents your engineers already know, one vendor's agent with hosting built in, or breadth.

## 8. What is the output?

A running app you can see and publish, a pull request in GitHub, or a chat reply. Founders and product teams usually want the app; engineering teams with a review process usually want the pull request.

## 9. How is it priced?

Per seat, per member who runs the agent, by credits, by sandbox hours, or flat per team. Compare on your real shape: two people who both prompt, five people where two prompt, or thirty engineers running agents in parallel produce very different bills from the same price list.

## How the tools answer

- **Replit, Lovable, Bolt:** parallel threads or drafts; review after; pooled or owner-paid credits; hosting included; own agents. Best for builders who want the whole platform. [Cofounder comparison](/multiplayer-ai-coding-workspace-for-startup-cofounders).
- **AQ, Coshell, MobSession, Omnigent:** live sessions for engineers; your own CLIs and keys; terminal or self-hosted. [Poly vs AQ](/poly-vs-aq), [Poly vs Coshell](/poly-vs-coshell).
- **Superconductor:** tickets and parallel implementations in hosted sandboxes; bring your own tokens; review at the pull request. [Poly vs Superconductor](/poly-vs-superconductor).
- **Claude Tag, hilos, Nairi, Bloome:** agents in the chat; requests and pull requests; org-level or flat pricing. [Poly vs Claude Tag](/poly-vs-claude-tag).
- **Poly:** live session; approval from any member before each change; receipt per turn per person; non-coders as equals; no keys; hosted sandboxes; Claude and ChatGPT; a running app; free to start then per person who runs the agent. Its limits are a desktop browser, static publishing and one turn at a time per room.

## Common questions

**Which question matters most?**

The first. Whether the team shares a live session, a project, or a chat decides what everyone will actually see and be able to do. The other eight follow from it.

**Is there a tool that answers all nine well?**

Not for every team. Poly answers them well for a small group steering one agent; Replit or Superconductor answer them better for many engineers working in parallel; Claude Tag answers them for a Slack-based organization on a Claude plan.

**How do I test a workspace before committing?**

Open it with one other person, give the agent a real task, and check three things: can the second person approve a change before it lands, does the turn say who spent what, and can either of you undo it. Most of the market fails at least one.

More guides: https://usepoly.co/guides · Security: https://usepoly.co/security · Pricing: https://usepoly.co/pricing