Poly vs Coshell for multiplayer AI coding

Coshell and Poly make the same claim, that a team should steer one AI coding agent together in one live session, and deliver it in opposite ways. Coshell is a self-hosted "drive" built on OpenCode: you run it on your own laptop or server for $0, bring your own model keys or Claude plan, and your team joins one shared session with a named prompt queue, plan approvals and a shared preview. Poly (usepoly.co) is hosted: nobody installs or configures anything, nobody holds a key, the AI can ask before each change, and every turn carries a receipt. Coshell is the choice for a technical team that wants to own its compute. Poly is the choice for a small group that wants to open a tab and build.

What Coshell does

Coshell, formerly ShellTab, started as a fork of OpenCode in May 2026. Its unit is the drive: a process on your machine, or on a server you deploy, that listens only on localhost and opens an outbound connection to Coshell's edge, so nothing listens on a public port. Drives run on a laptop, through one-click Render or Railway templates, on Northflank (including your own AWS, Google Cloud or Azure account), or from a cloud-init script on DigitalOcean, Hetzner, Vultr or Linode.

Inside a drive, one live session holds the whole team:

• A named prompt queue. Prompts wait in order with a name on each, and since September "Send now" lets anyone steer a running turn, including a teammate's, with the steer attributed.
• Follow mode. Follow a teammate's session, file and cursor.
• Plans first. A Plan tab where the team approves a plan, commenting on any line, before the agent executes it. Admins choose whether the agent's edits are accepted automatically, and can lock that for the whole drive.
• Shared browser tabs. Every dev server in the drive shows up as a tab everyone can open.
• Attributed commits. Each commit is authored to whoever prompted it.
• Overlap. When a second agent edits the same lines, Coshell holds the edit and arbitrates. It doesn't merge code, and it fails open.

Models come from your side: Claude Pro or Max through Claude Code, OpenAI models on your own keys, and by Coshell's count more than 70 models through OpenCode, free ones included. You reach a session from the browser, a terminal client (coshell attach), Slack, a phone, or a grid of several sessions at once, and an @coshell assistant answers in the session chat.

Recent changes, from Coshell's changelog: projects no longer need git, you can fork a session from any message (both September), and October 23 brought Team Ask, Code Watch, PR Watch and messages between agents in different sessions.

The same evening in each

Say three friends are building a side project after work.

With Coshell, one of them installs a drive on a spare server, connects their Claude Max plan, and shares an invite. The others join from a browser or a terminal. They queue prompts, approve the agent's plan line by line, and open the dev server as a shared tab. The model spend is on the host's Claude plan or API keys, and if the server is off, the drive is off.

With Poly, one of them creates a room and invites the other two by email. Nobody installs anything or connects a plan; each person's turns draw on their own free credits. They pick a trust level, say "Ask before big steps", and every proposed change appears as a card any of them can approve. The room keeps going whether or not anyone's laptop is open, and the group chat stays between the three of them.

Where Coshell is stronger

• Free on your own compute. No seats, no metering, your keys, your machine.
• Model breadth. Dozens of models through OpenCode, including free and self-hosted ones.
• Surfaces. A terminal client, Slack, a phone view and a grid of sessions, alongside the browser.
• Steering mid-turn. "Send now" redirects a running turn. In Poly, a new prompt waits for the current turn to finish, or someone presses Stop.
• Overlap. A referee for two agents editing the same lines, which Poly doesn't need because a room runs one turn at a time.
• Enterprise options. SSO, SCIM, on-prem deployment, models routed through Bedrock in your own AWS account, and HIPAA with a business associate agreement.

Where Poly is stronger

• Nothing to run. Coshell's first step is installing and hosting a drive, or a cloud deploy, and keeping it up. Poly's first step is a browser tab, and every account starts with free credits and no card.
• Approval per change, from any seat. Coshell approves plans, and its admins decide whether edits are accepted automatically. Poly's trust levels can put each file change and command on every member's screen before it happens, and any member can answer.
• Receipts. Poly names who ran each turn, which model and what it cost, charged to that person. Coshell attributes commits to the prompter, while the spend lands on whoever's keys or plan the drive uses.
• No keys anywhere. Coshell runs on your keys, on a machine your teammates share. Poly's platform keys never enter a sandbox, and members never hold one.
• A channel the AI cannot read. Coshell's session chat includes an assistant that answers in it. Poly's group chat and polls are hidden from the agent by design.
• Built for people who don't code. Coshell's pages speak to engineering managers, startups, remote teams and agencies. Poly's room is designed so the person who can't code can prompt, approve and undo.

Security and where your code runs

On Coshell your code, your keys and your agent all live on the drive you host, and Coshell's edge only relays an authenticated outbound tunnel. That puts you in charge of the machine, its updates and who can reach it. Every plan has an audit trail tied to verified identities, and deleted data is permanently removed after 30 days.

On Poly, each turn runs in a fresh sandbox under gVisor on Poly's servers, with a read-only root filesystem, no direct internet access and no API keys inside; a domain outside a short allowlist waits for a member to allow it. The security page has the detail, and secure shared coding-agent workspaces covers what changes once several people share one agent.

Which should you pick?

• A technical team that wants to self-host on its own compute with its own keys, and is comfortable operating it: Coshell.
• You want many models, including free and local ones: Coshell.
• A small group, including non-engineers, that wants to open a browser and build, with approvals and receipts: Poly.
• A group that wants the argument about the plan to stay out of the prompt: Poly.

If you're weighing OpenCode itself, OpenCode vs Claude Code compares the two agents. More shared-session tools are in Collaborative Claude Code tools, and another bring-your-own-subscription workspace is compared in Poly vs Tutti VM.