Rush runs agent work unattended and hands back the evidence. You describe a job once; an agent (Claude Code, Codex or OpenCode) does it in its own cloud computer, with a desktop and a browser, and every run comes back with what it did, so you can check it and run it again. The API, the MCP server and the console at cloud.getrush.ai all start the same runs.

Which agents does it run?

Three: Claude Code, OpenAI Codex and OpenCode. You pick one per run, and the run records which harness and which version did the work.

You send the task in plain words, the agent and, when the work touches code, a repository; in the console you can also set it to run on a schedule, a webhook, a GitHub event, a Linear ticket or an email. Each run starts in a fresh cloud machine shared with no other run. You can follow it live, send a follow-up or cancel it, and it comes back with the change, the files it made and every step the agent took, flagged when it finished without the proof you asked for.

Teams with work they would rather hand to an agent every time it comes up: reviewing pull requests, fixing a failed build, triaging tickets, keeping docs current, pulling a report together. If you already use a coding agent, Rush is where it keeps working without you, with a record you can review before anything ships.

Free includes 5 agent-hours a month with no card. Pro is $20 a month for 40 agent-hours and 4 agents at once. Team is $200 a month for 400 agent-hours and 12 agents at once. You pay for the cloud the agent runs in, not for seats.

Do I need my own model account?

No. Rush supplies the models: Claude, GPT, Kimi, DeepSeek, Grok and Muse Spark. Your first $1 of models is on us; after that, tokens are billed at the provider's cost plus 20% on a paid plan with a spend limit. Every run started with the API uses them, so the agents you build for your customers need no account of theirs or yours.

Can I use my own model subscription or API key?

Yes, for your own runs in the console: they use your Claude Code or Codex seat, or your own API key, and the model stays between you and your provider. An API run uses one only when it names it. The run still happens on Rush hardware, so it draws on your plan's agent-hours.

Rush: Leave it running.
Check what it did.

Rush runs agent work unattended. Describe a job once; an agent does it on its own cloud computer and browser, and every run brings back the change, the files and each step. Check it, adjust it, run it again.

rush run claude-code -r owner/repo "fix the flaky test"

Start a run

Free to start. No card.

A living room at cruising altitude: two children build a tower of wooden blocks on the rug beside a sleeping golden retriever while their mother reads on the sofa, under a round window framing a wing above the clouds.

A real run

What came back.

One request asked Claude Code to review pull request #2500 on our public repository. It came back with 43 turns and 42 tool calls, every one on the record.

from request to done: 7 min 32 s
events on the record: 187
for the whole run: $0.23

cloud.getrush.ai · run-af7b2434

Rush console showing a completed Claude Code run that reviewed pull request #2500 on muqsitnawaz/agents: 187 events, 43 turns, 42 tool calls, 1,544,815 tokens, $0.23, with the Environment panel listing status, agent, model, repository and pull request.

Captured from the Rush console on September 10, 2026.

How it works

Describe it once. Check every run.

01
Describe the job
In plain words, with one API call or in the console, where it can also run on a schedule, a webhook, a GitHub event, a Linear ticket or an email.
1. Request
2. Its computer
3. Live work
4. Record
POST /o/{org}/p/{project}/sessions
```
$ rush run claude-code \
    -r owner/repo \
    "fix the flaky test" \
    --keep
```
the harnessyour repositorythe task, in plain wordskeep it, so you can write to it
or the same request to the API{ "agent": "claude", "prompt": "fix the flaky test", "persistence": "resumable" }
202 Accepted{ "id": "ag-claude-1a2b3c4d", "status": "queued" }
Fig. 1 · The request, from the CLI or the API
02
It gets its own computer
A fresh cloud machine with a desktop, a browser and your project's context, shared with no other run.
1. Request
2. Its computer
3. Live work
4. Record
GET …/sessions/ag-claude-1a2b3c4d
queuedallocatingrunning
Its own computer
machine_state: running
repository
owner/repo, checked out
context
your project's skills and plugins
harness
claude-code, its version on the record
persistence
resumable (--keep)
Shared with no other run. Nothing runs on your laptop.
Fig. 2 · The session's status and the machine it runs on
03
Watch, steer or walk away
Follow it live, send a follow-up, or take the controls of its screen. Close the laptop; nothing waits on it.
1. Request
2. Its computer
3. Live work
4. Record
GET …/sessions/ag-claude-1a2b3c4d/events · text/event-stream
event: thinkingThe retry test races its timer.
event: tool_usebash · bun test retry.test.ts
event: tool_result1 failed, 11 passed
event: tool_useedit · retry.test.ts
event: tool_result12 passed
event: messageThe test now waits on the timer instead of sleeping.
```
$ rush send ag-claude-1a2b3c4d "cover the retry path too"
```
202 · status: queued, position: 1 · delivered when the turn ends
Fig. 3 · The live stream, and a follow-up while it runs
04
Check it, then make it better
Each run comes back with the change, the files and every step, flagged when it finished without the proof you asked for. Adjust the job and run it again.
1. Request
2. Its computer
3. Live work
4. Record
event: done · GET …/sessions/ag-claude-1a2b3c4d/result
completed
The change
branch
rush/fix-flaky-test
pr_url
github.com/owner/repo/pull/…
result
The test waits on the timer; retry path covered.
The record
…/trajectory
every message, tool call and result
total_cost_usd
tokens and cost, on the session
machine_state: pausedKept for 7 days, $0 while parked, until you write to it again.
Fig. 4 · What comes back when it is done

How a run works, end to end

Computers and browsers

A real desktop. A fast browser.

Every run has a desktop and a browser of its own, and you can watch its screen or take the controls. We also build browser and computer tools for agents, and test them against Browser Use and the Cua MCP server.

2.2×

faster than Browser Use on Browser Use Bench tasks, at 7% lower cost per task

browser-cli · 4 tasks, 16 runs per tool · Oct 19, 2026

14.4 s

per task on eight live sites, against 16.8 s for Browser Use, every answer correct

browser-cli · 8 tasks, 2 runs each · Oct 19, 2026

16%

lower cost per run than the Cua MCP server on 20 hard OSWorld tasks, passing 51 of 58 runs to its 49

computer-cli · OSWorld-Verified, 3 trials each · Oct 20, 2026

We ran these ourselves with Claude Sonnet 5.5 in Claude Code, on small task sets. Every score and how it was measured are in the OSWorld write-up and the browser-cli README.

A father and two children sit on the edge of a travertine terrace above a sea of clouds at dusk, pointing at the sunset, the lit living room behind them.

Persistent agents

Home before dark.

Close the laptop and the work keeps going. When a persistent agent goes quiet it pauses with its memory and files intact, costs nothing while it is parked, and picks up where it left off the moment you write back.

How runs pause and resume

Shared work

Work your team can see.

A run doesn’t stay in one person’s terminal. Your team can read it, comment on what came back and start runs of their own.

Shared projects

Share a project with your org. Everyone sees its runs and what came back, comments on it, and starts runs of their own.

Shared agents

Share a saved agent with your team, or with the teammates you name. Each of them runs it with their own accounts.

Pages, images and video

Publish a report, images or a screen recording as a page, with a transcript for video. Keep it to your org, or share it with people by email.

Pricing

Pay for the hours, not the seats.

You pay for the cloud the agent runs in, measured in agent-hours. No seats, no per-agent subscriptions, no credits to pre-buy.

Free

No card

$0forever

5 agent-hours a month

Connect a repository and run an agent in the cloud today.

Pro

One developer

$20/month

40 agent-hours a month

Covers a working developer's month, four agents at a time.

Team

One developer, many agents

$200/month

400 agent-hours a month

Ten times Pro's hours, twelve agents in parallel.

Bring your Claude Code or Codex seat, or your own API key, and the model stays between you and your provider. The run still happens on Rush hardware, so it draws on your plan’s agent-hours.

Compare plans

Questions

The short answers.

: Rush runs agent work unattended and hands back the evidence. You describe a job once; an agent (Claude Code, Codex or OpenCode) does it in its own cloud computer, with a desktop and a browser, and every run comes back with what it did, so you can check it and run it again. The API, the MCP server and the console at cloud.getrush.ai all start the same runs.
: Three: Claude Code, OpenAI Codex and OpenCode. You pick one per run, and the run records which harness and which version did the work.
: You send the task in plain words, the agent and, when the work touches code, a repository; in the console you can also set it to run on a schedule, a webhook, a GitHub event, a Linear ticket or an email. Each run starts in a fresh cloud machine shared with no other run. You can follow it live, send a follow-up or cancel it, and it comes back with the change, the files it made and every step the agent took, flagged when it finished without the proof you asked for.
: Teams with work they would rather hand to an agent every time it comes up: reviewing pull requests, fixing a failed build, triaging tickets, keeping docs current, pulling a report together. If you already use a coding agent, Rush is where it keeps working without you, with a record you can review before anything ships.
: Free includes 5 agent-hours a month with no card. Pro is $20 a month for 40 agent-hours and 4 agents at once. Team is $200 a month for 400 agent-hours and 12 agents at once. You pay for the cloud the agent runs in, not for seats.
: No. Rush supplies the models: Claude, GPT, Kimi, DeepSeek, Grok and Muse Spark. Your first $1 of models is on us; after that, tokens are billed at the provider's cost plus 20% on a paid plan with a spend limit. Every run started with the API uses them, so the agents you build for your customers need no account of theirs or yours.
: Yes, for your own runs in the console: they use your Claude Code or Codex seat, or your own API key, and the model stays between you and your provider. An API run uses one only when it names it. The run still happens on Rush hardware, so it draws on your plan's agent-hours.

From the blog

Notes from the work.

Oct 20, 2026

Imagine what you’ll do with the time.

Describe it once, check every run, and let it keep going. Free to start, no card.

Start a run Read the docs

Rush: Leave it running.
Check what it did.

What came back.

Describe it once. Check every run.

Describe the job

It gets its own computer

Watch, steer or walk away

Check it, then make it better

A real desktop. A fast browser.

Home before dark.

Work your team can see.

Shared projects

Shared agents

Pages, images and video

Pay for the hours, not the seats.

Free

Pro

Team

The short answers.

Notes from the work.

Your Agent Can Use Any App. For a Quarter Fewer Tokens.

Agents Now Write Their Own Instructions. Who Reviews Them?

Our Writing Agent Made 858 Commits. Then It Died Quietly.

Imagine what you’ll do with the time.

Rush: Leave it running.Check what it did.

What came back.

Describe it once. Check every run.

Describe the job

It gets its own computer

Watch, steer or walk away

Check it, then make it better

A real desktop. A fast browser.

Home before dark.

Work your team can see.

Shared projects

Shared agents

Pages, images and video

Pay for the hours, not the seats.

Free

Pro

Team

The short answers.

Notes from the work.

Your Agent Can Use Any App. For a Quarter Fewer Tokens.

Agents Now Write Their Own Instructions. Who Reviews Them?

Our Writing Agent Made 858 Commits. Then It Died Quietly.

Imagine what you’ll do with the time.

Rush: Leave it running.
Check what it did.