Poly vs AQ for running coding agents as a team

AQ and Poly both let a team join one live coding-agent session, and they are built for different people. AQ runs real agent CLIs such as Claude Code and Codex on a dedicated machine, in your own cloud if you want, and teammates join the same live terminal, editor and app preview. Poly (usepoly.co) runs one AI (Claude, GPT, Grok or Muse) in a browser room where every member can prompt, approve a change before it happens, and undo any turn, with a receipt on each. AQ is for engineering teams who already live in the CLI. Poly is for small groups where at least one person doesn't.

What AQ does

AQ, from Betterleap Inc., pitches itself as "Run every coding agent in the cloud, with your team." It is in early access: you sign up with a work Google account and wait in a review queue, usually a day or two.

• Workspaces. Each workspace is an isolated git worktree on a dedicated VM. The free sandbox sleeps when idle and wakes in about 30 seconds; on Pro the machine never sleeps, and it can live in your own cloud account.
• Any agent. Claude Code, Codex, Cursor, Devin, Grok, Kimi, Antigravity, Muse Code, OpenCode, Pi and others run under each user's own CLI login, with no markup, plus AQ's built-in agent, billed per token.
• Shared sessions. Teammates open a session by link and see the same terminal, editor and running app. The session's owner can hand terminal control to a teammate for a few minutes. On Pro, reviewers pin comments on elements of the running app and send them to the agent as prompts.
• Engineering integrations. Labelling a Linear issue ai-task prepares a workspace for it; no agent starts until a person launches one. Pro adds a REST API, signed webhooks and an MCP server.
• New in October. Hinoki Offensive One, a security model that reads a diff or a whole codebase looking for vulnerabilities, paid from AQ credits (October 25).

The same task in each

Say a three-person startup needs a settings page. One engineer, one designer, one founder who doesn't code.

In AQ, the engineer opens a workspace and starts Claude Code in its terminal. The designer and the founder open the session link and watch the terminal scroll and the app preview update. The designer pins a comment on a button in the preview ("make this the brand blue"), which goes to the agent as a prompt. If the founder wants to type, the engineer hands over the terminal. The agent runs on the engineer's Claude subscription, and nothing pauses for approval: AQ assumes a human is steering.

In Poly, any of the three types the first prompt into the room. If the room is set to "Ask before big steps", the AI's plan to add a database table shows up as a card on all three screens, and the founder can approve it as easily as the engineer. Each turn ends with a receipt saying whose credits paid for it. The designer and the founder argue about the blue in the group chat, which the AI doesn't read, and then one of them sends the decision as the next prompt.

Where AQ is stronger

• Real terminals, real CLIs, your own cloud. If your team's workflow is Claude Code in a terminal, AQ is that terminal, shared, with full custody of the machine on Pro.
• Any agent. More than ten CLIs, each on the user's own subscription with no markup.
• Engineering integrations. Linear, a REST API, webhooks and an MCP server.
• Comment on the running app. Pinned comments that become prompts are a good idea, and AQ shipped them.
• Parallel workspaces. One worktree per task, so a team can run several at once. A Poly room's shared agent runs one turn at a time.

Where Poly is stronger

• A say before it acts. AQ has no approval step for the agent's actions; what it does have is a handoff of terminal control. Poly's trust levels can put each proposed file change and command on every member's screen, in plain language, before it happens.
• Receipts. Every Poly turn names who ran it, which model, how long it took and what it cost, charged to that person. AQ bills per seat, and the agent spend sits on each user's own subscription.
• No keys, no CLI logins, no waitlist. Nobody in a Poly room configures anything, and anyone can sign up and start on free credits without a card. AQ users sign into each CLI themselves, after an early-access review.
• Built for the non-engineer. AQ names product managers and designers, but its surface is a terminal. Poly's is a timeline, cards and a preview, on a computer or a phone.
• A channel the AI cannot read. Poly's group chat and polls are hidden from the agent.

Data and where your code runs

AQ runs your code on its dedicated machines or, on Pro, in your own cloud account. Its terms, updated October 9, 2026, let it use customer data to train its own models. Paid plans can opt out; the free plan can't. AQ publishes no SOC 2 report.

Poly runs each turn in a throwaway sandbox on its own servers, with no direct internet access, no API keys inside and a human answer needed for any domain outside a short allowlist. Turns go to the lab behind the chosen model under Poly's own accounts. Poly can't run in your cloud. The details are on the security page, and secure shared coding-agent workspaces compares the approaches.

Pricing, side by side

AQ is free for a personal sandbox. Pro is $50 per user per month at the early-access rate, kept for your first 12 months, then $100, against a listed $200. Seats are billed only once someone joins, and each includes $10 of AQ credits a month. The CLIs themselves run on each person's own subscriptions.

Poly is free to start with no card. Whoever runs the AI pays $20 a month for Pro or $50 for Max, and everyone else watches, chats and approves for free. A five-person team with two people prompting pays Poly $40 a month and AQ $250 at the early-access rate, plus whatever Claude or ChatGPT plans its engineers already pay for.

Which should you pick?

• Engineers who want a shared cloud terminal for the CLIs they already use, with Linear and GitHub wired in: AQ.
• A team that wants to keep the agent's machine in its own cloud account: AQ Pro.
• A small group including people who don't code, who want to approve each change and see who spent what: Poly.
• You want Claude, GPT, Grok and Muse in one room without anyone holding a subscription: Poly.

More shared-session tools are in Collaborative Claude Code tools and Multiplayer AI tools. For other agent workspaces built around engineering teams, see Poly vs Superconductor and Poly vs Conductor.