Cursor is an excellent AI editor for one person, and since March its cloud agents can run their tool calls on machines you own. Coshell is a different thing: the shared live session your whole team and its agents work in together, with the whole agent running on a machine you control. Here is the honest comparison.
Cursor is one of the best AI editors there is, and we are not trying to beat it.
Coshell is not an editor. It is the shared session the whole team works in, with the agent, at the same time, on machines you control.
Overview is in with the three stat cards and the recent orders table. Live on http://localhost:5173.
Cards stretch now. The grid uses align-items: stretch and each delta line is pinned to the bottom with margin-top: auto.
Added the toggle. 30 days aggregates by week and the axis relabels. 8 tests pass.
Delta sits on the top row now, right aligned, so the numbers line up down the column.
Token bucket middleware is in with 12 tests. Step 1 verified.
Moved the bucket to Redis with a Lua script so the check is atomic across replicas. 14 passing. Steps 2 and 3 verified.
Webhooks now share the limiter. 16 passing. Marking step 4 verified and moving to the flag rollout.
Node.js isn’t installed on this VM, so I can’t start the dev server yet. I can install it with the system package manager and then run it. Want me to?
Understood, Node 22 LTS from NodeSource.
Server is running at http://localhost:5173/, Vite dev server ready in 284ms.
Eight commits since main: the coupon model, the checkout hook, and a migration. No PR yet.
No. You have no PLAN, TODO, or roadmap file.
Your closest thing to a plan is .coshell/review/dod.md: integer cents, handled promises, no focused tests, no dead exports.
Drafted docs/plans/discounts.md with five steps. It is waiting for approval in the Plan tab.
Hi. Big Pickle here, ready when you are.
2, 3, 5, 7, 11, 13, 17, 19, 23, 29.
1
2
3
4
5
6
7
8
9
10
11
12
bun test after every step, load test before the flag movesProtect /v2 from bursty API keys without slowing normal traffic. One key should never be able to starve the others.
A token bucket per API key, 100 requests per minute, stored in Redis so all three replicas share state. Over-limit requests get a 429 with a Retry-After header. Ship behind the rate_limit_v2 flag and roll out gradually.
bun test after every step. Load test with 1k keys before the flag moves to 50%. Watch 429 rate and p99 in the dashboard for a day at 10%.
Overview is in with the three stat cards and the recent orders table. Live on http://localhost:5173.
Cards stretch now. The grid uses align-items: stretch and each delta line is pinned to the bottom with margin-top: auto.
Added the toggle. 30 days aggregates by week and the axis relabels. 8 tests pass.
Delta sits on the top row now, right aligned, so the numbers line up down the column.
Token bucket middleware is in with 12 tests. Step 1 verified.
Moved the bucket to Redis with a Lua script so the check is atomic across replicas. 14 passing. Steps 2 and 3 verified.
Webhooks now share the limiter. 16 passing. Marking step 4 verified and moving to the flag rollout.
Node.js isn’t installed on this VM, so I can’t start the dev server yet. I can install it with the system package manager and then run it. Want me to?
Understood, Node 22 LTS from NodeSource.
Server is running at http://localhost:5173/, Vite dev server ready in 284ms.
Eight commits since main: the coupon model, the checkout hook, and a migration. No PR yet.
No. You have no PLAN, TODO, or roadmap file.
Your closest thing to a plan is .coshell/review/dod.md: integer cents, handled promises, no focused tests, no dead exports.
Drafted docs/plans/discounts.md with five steps. It is waiting for approval in the Plan tab.
Hi. Big Pickle here, ready when you are.
2, 3, 5, 7, 11, 13, 17, 19, 23, 29.
1
2
3
4
5
6
7
8
9
10
11
12
| Cursor | Coshell | |
|---|---|---|
| What it is | An AI editor on your machine, plus cloud agents that can run their tool calls on machines you own | A shared live session your team and its agents work in, on a drive you host |
| Single-player or shared | Single-player: one copy per person in the editor, one worker per cloud agent run | Shared: one live session for the whole team, or private when you want it |
| Real-time multiplayer | Teammates can open a cloud agent run by link, read-only unless admins enable follow-ups | Several people prompt the same agent in one session, at the same time, with a name on every prompt |
| Overlap detection | Separate runs land as separate pull requests; you reconcile in review | An overlap layer catches the same concern being changed twice, while it happens |
| Load the codebase once | Each person's editor indexes its own copy | The shared session loads context once, for everyone |
| What runs on your machine | The editor. For cloud agents, the tool calls: file edits, terminal commands, local MCP, computer use. The agent loop stays in Cursor's cloud | The whole agent runtime, with the files, the terminal, and the preview |
| Where inference runs | Cursor's cloud, on Cursor's Composer models and frontier models through Cursor | Your provider, under your key, or your Claude Pro or Max subscription |
| What leaves your network | For cloud agents: file contents, terminal output, diffs, screenshots, and MCP results, sent to Cursor (per their docs) | Prompts to your model provider. The Coshell edge relays the browser session and never sees a model call |
| Pools and autoscaling | Named pools, a controller, autoscaling, hibernate and restore (ahead of ours; Enterprise plan) | Not a pool. One drive per project or workstream, with as many sessions as you want |
| Computer use | Click, type, screenshot, and drive a browser on a Linux or Mac worker's own desktop; watch or take over | A cloud browser (Notte, on your key) the agent drives and you can take over and hand back, plus live preview of localhost in the session |
| Per-session privacy | The editor is local. Cloud runs are visible to the Cursor team they started under; Privacy Mode keeps code out of training | Per session: private, named teammates, or whole drive |
| Models | Built-in models, plus your own keys in the editor; cloud agents billed through your Cursor plan | 70+ models on your own keys, including a Claude Pro or Max subscription |
| Where you work from | The desktop app, web, iOS app, Slack, GitHub, Linear, API | Any browser, the CLI, VS Code, or your phone |
| Git | Commits from your local checkout; cloud agents push a branch and open a pull request | Commits by the prompter, pull requests as the connected person, no token on the drive |
| Self-hosting | Tool execution on your machines; the agent and inference stay in Cursor's cloud. Pools need Enterprise | The whole agent on your laptop, VM, or container, free; Enterprise adds org-wide on-prem |
| Pricing | Hobby free; Pro from $20 a month per person; Teams $40 per user a month; pools on Enterprise | Free to self-host; Enterprise is paid |
| Best for | One developer moving fast in their own editor, with cloud agents that need your VPN, caches, or GPUs | A team working with its agents in the same live session, on its own machines |
Not exactly. Cursor is an AI editor that runs on your machine for one person, and its cloud agents are good at long tasks that end in a pull request. Coshell is a shared session your whole team joins to work with the same agents at the same time, on a drive you host. Many teams use both: Cursor for solo editing, Coshell when the work is shared. If you searched Cursor vs coshell expecting two editors, that is the key difference: one is an editor, the other is a shared session.
In the editor, no: each person edits their own copy of the code on their own machine. For cloud agents, partly: per Cursor's docs a teammate can open a run by its link to read the conversation and the changes, and a team admin can turn on follow-ups so teammates can send messages into it, but each chat is claimed by one worker. Coshell starts from the other end. Several people prompt the same agent in one live session at the same time, every prompt and commit carries who sent it, and you can keep a session private when you want to.
No. They self-host different layers. With Cursor, your machine runs the tool calls: file edits, terminal commands, local MCP, computer use. Cursor's docs say the agent loop, inference, and planning run in Cursor's cloud, and that during a run the worker sends file contents, terminal output, diffs, and screenshots to Cursor. With Coshell, the agent itself runs on your machine and calls your model provider with your key. Nothing about the run depends on our cloud except signing your team in and relaying the session to their browsers.
The model call never does. Your drive sends prompts straight to the provider you chose, under your key or your Claude subscription. What passes through Coshell is the session itself: the drive keeps an outbound, encrypted tunnel to our edge, which signs your teammates in and relays what they see in the browser. The Enterprise tier runs that edge on your own infrastructure too, so nothing leaves your environment.
No. Two things prevent it. Unlike each person editing their own checkout, everyone shares one running copy, so there is no divergent branch to lose in a later merge. And when two efforts reach for the same concern, the overlap layer catches it and steers the second one before it overwrites the first.
Fleet scheduling. A pool is a named queue of workers with a controller that claims requests, spawns machines on demand, autoscales on queue depth, and hibernates idle machines until the next prompt, on Kubernetes, Lambda, Modal, E2B, and others. Mentions in Slack, GitHub, and Linear route into a pool by name. If you want a fleet of short-lived workers, one per task, Cursor's pools are built for that. A Coshell drive is the opposite shape: one persistent place per project that the whole team keeps, with as many sessions in it as you want.
Yes. Coshell runs on 70+ models with your own keys, and a drive runs the Claude Code binary on your Claude Pro or Max login if you have one. Work pushes to your real git repositories as a normal git operation. Self-host a drive on your own machine and the agent, the files, and the model calls all stay under your control; the Enterprise tier adds org-wide on-prem and SSO. Cursor's cloud agents run on the models Cursor offers, billed through your Cursor plan.
Nothing. Install the CLI, run coshell host init on a laptop, a VM, or a container, or deploy with one click on Railway, Render, or Northflank, and the whole team joins the same session in a browser. You bring the compute, so there is nothing for us to meter. Enterprise is the paid tier: SSO, an audit trail, and on-prem at scale.
Keep Cursor for solo work. When the work is shared, self-host a drive for free and put the whole team in one live session with the agent.