Poly vs Superconductor for teams and coding agents

Superconductor and Poly both call themselves multiplayer, and they organize work differently. Superconductor is a workspace for a team with a backlog: a ticket spawns one or several agent "implementations", each in its own cloud sandbox, and the whole workspace can chat in any of them, check the result with a QA agent, and ship a pull request. Poly (usepoly.co) is a room: one AI, one timeline, every member prompting, approving a change before it happens, and undoing any turn, with a receipt on each. Superconductor has far more surface area: nine agent harnesses, native desktop and mobile apps, benchmarks and dozens of integrations. Poly is the narrower tool for a pair or a small group, including people who don't code, with no keys to bring.

What Superconductor does

Superconductor, made by Volition, Inc., runs coding agents in isolated cloud sandboxes, with configurable network access and Admin, Collaborator and Viewer roles. It's a different product from Conductor, which turns up in the same searches; Poly vs Conductor covers that one.

• Nine harnesses. Claude Code, Codex, Grok Build, Cursor, Pi, Muse Code, OpenCode, Amp and Factory Droid. Muse Code, from Meta, was added on October 19.
• Tickets and implementations. The unit of work is a ticket. An implementation is one agent in one sandbox working on it; several can run side by side, and a recommendation picks the best. Assigning a Linear issue launches an agent.
• Review. Every implementation has a live preview and artifacts, QA checks (in beta) look at the work before review, and guided review supports inline comments.
• Apps and integrations. Native apps for macOS and Linux (Windows is "coming soon"), iPhone and iPad, and an Android web app; Slack and GitHub; email and meeting agents that turn messages and transcripts into tickets; and more than forty integrations, Linear, Notion, Figma, Sentry and Jira among them.
• Beyond code. Since early October a project doesn't need a repository, and Docker is optional. Superconductor now pitches research, sales, marketing and support work as well as engineering.

Models are mostly bring-your-own: connect a Claude, ChatGPT, SuperGrok or Muse plan, or API keys, though Claude Code can also run on the workspace's credits. Sandboxes run on Modal and Runloop, and the platform is on AWS in the United States. Superconductor says it doesn't train on your code; its SOC 2 Type 2 report is in progress with no date.

How the team talks to the agent

In Superconductor, every message in an implementation's chat reaches the agent, labelled with the sender's name. Since September, a message that starts by @mentioning a person is still read by the agent, but the agent skips its turn. The docs say Claude Code's AskUserQuestion tool isn't supported, and there is no approval step for the agent's actions: the sandbox's isolation and the review before merge do that job. When the agent hits a domain the network policy blocks, someone chooses "Allow for project" or "Allow for this ticket", and the agent doesn't resume on its own.

In Poly, the group chat and polls sit beside the timeline and the AI never reads them, so people can disagree about the plan without the disagreement becoming a prompt. The AI asks questions on cards anyone can answer, and at the trust levels that ask ("Show a plan first", "Ask before big steps", "Ask about everything") each proposed change waits for a member's yes. Blocked domains work much like Superconductor's: the connection is held, a card names the domain, and any member can allow or deny it.

Where Superconductor is stronger

• Breadth. Nine harnesses, native apps on desktop and phone, benchmarks on your own repo (Pro), QA checks and guided review, and dozens of integrations.
• Parallelism. Several implementations of one ticket at once, compared side by side.
• An engineering workflow. Tickets, review, pull requests and automations for a team with a backlog, now with work that needs no repository.
• Pace. A weekly changelog through September and October 2026, with new models and connectors most weeks.

Where Poly is stronger

• No keys or plans to connect. Most of Superconductor's harnesses run on a plan or key someone connects; a non-technical cofounder has to sort that out before doing much. In Poly nobody holds a key, usage is billed per person as credits, and every account starts free.
• Receipts. When spend sits on each user's own subscription, the workspace can't show who ran what for how much. Poly's receipt on every turn does exactly that.
• A say before it acts. Superconductor relies on isolation plus review at the pull request. Poly can put each proposed change on every member's screen first.
• The room, not the ticket queue. Poly organizes around one conversation the agent is part of. For a pair building one thing that is the whole product; for an engineering org it may read as missing structure.
• A channel the AI cannot read. Every message in a Superconductor implementation reaches the agent. Poly's group chat never does.

Pricing, side by side

Superconductor meters usage in credits, which replaced sandbox hours in October. Free starts with 20,000 credits (which it values at $50, roughly 200 sandbox hours at one vCPU), up to four members and four agents at once. Pro is $100 a month, with credit tiers up to 400,000, overage at $25 per 10,000 credits, and up to 40 members. Enterprise is priced on request and can run sandboxes in your own cloud. Model tokens are on top, on your own plans or keys.

Poly is free to start with no card. Whoever runs the AI pays $20 a month for Pro or $50 for Max, and watching, chatting and approving are free. The models are included in the credits.

Which should you pick?

• A team with a backlog that wants parallel implementations, QA, review and native apps, and has its own Claude or ChatGPT subscriptions: Superconductor.
• A pair or small group, some of whom don't code, who want to build one thing together with no keys and a receipt on every turn: Poly.
• You want the agent's work as tickets and pull requests: Superconductor. You want it as a conversation with an Undo button: Poly.

For a larger engineering team, multiplayer AI coding for a 50-person engineering team compares the options at that size, and AI code review tools covers the review side. A terminal-first alternative is compared in Poly vs AQ.