Compare · Agent sandboxes · verified 17 Oct 2026
Rush vs Modal Sandboxes
Modal Sandboxes are one product on a GPU and serverless platform. Rush is one job: an agent doing a task and returning it with proof.
Facts checked 17 October 2026. Prices and limits move. Read the citations.
Pick Modal when you need GPUs or RL-scale fan-out. Pick Rush when an agent should do a task on a repository and return the pull request.
Proof
A real Rush run
Console capture of a coding Task that opened a pull request. Not a mockup.
Read this first
The useful distinction
Modal Sandboxes: Programmable gVisor containers, or a VM runtime, on Modal's serverless cloud, for agent code, RL rollouts and untrusted code, with snapshots, egress rules and GPUs.
A sandbox is a computer for an agent. You still write the agent loop, clone the repository, push the branch and decide when the work is done. Rush is the layer above: you send the task and get back the pull request, report or files.
Head to head
Buyer questions
| Question | Rush | Modal Sandboxes |
|---|---|---|
| What runs | Claude Code, Codex or OpenCode, the vendors' own CLIs, or a saved agent built on one. | None built in. You bring the agent and call Sandbox.create from Python. |
| Isolation | One E2B microVM per run with its own disk, memory, desktop and browser. It parks between turns and is removed on cancel or after 7 idle days. | gVisor by default, or a VM runtime with its own kernel. Default lifetime is 5 minutes and the timeout goes up to 24 hours; past that you restore from a filesystem snapshot. |
| Result and pull request | A run on a repository works on its own branch and opens the pull request; a run whose checks fail ends needs_review, not completed. Reports and files come back as artifacts. | Left to you. The sandbox guide has no git or pull request feature. |
| MCP | Rush is an MCP server at api.prix.dev/mcp with 18 tools. A run also takes MCP servers by connection or inline. | Not mentioned on Modal's sandbox guide. |
| Team | Everyone in the org sees the same Projects and runs, with who started each. Automations start runs from GitHub events, schedules, webhooks and Linear. | Workspaces, environments, service users, user groups and role-based access control. |
| Starting price | Free 5 agent-hours / month; Pro $20; Team $200 | Starter $0 + compute with $30 / month credit; Team $250 / month + compute |
| Where it taps out | Three harnesses. Repositories through the GitHub App only. No self-host or customer VPC. A key runs 4 at once by default, up to 100. | 5-minute default lifetime and a 24-hour cap per sandbox. Sandbox CPU costs about three times a Modal function's. GPU sandboxes need the gVisor runtime. |
Decision
When each side wins
Rush wins when
- The job is a coding or research task that should end in a pull request, report or files.
- You want the agent installed and run for you instead of writing the loop.
- A run should park between turns at no charge and wake with its memory.
Modal Sandboxes wins when
- You need GPUs inside the sandbox, billed per second.
- You fan out at RL scale: Modal advertises 1M+ sandboxes in under a minute.
- You want CIDR and domain egress allowlists and one platform for inference, training and sandboxes.
Rush
Where Rush sits
Rush does not compete with Modal on GPUs or RL rollouts. It covers the coding-agent case Modal lists as a use case, with the agent, the clone, the pull request and the transcript already built.
Honest limits
Skip Rush if
- Rush runs three agents: Claude Code, Codex and OpenCode. It does not host an agent you wrote yourself.
- Repositories come through the Rush GitHub App. A GitLab or Bitbucket repository cannot be a run's repository yet.
- There is no self-hosted or customer-VPC deployment, and we do not publish a compliance report today.
- Compute is metered. Free includes 5 agent-hours a month and Pro 40, each with a weekly ceiling.
- You review the pull request. Rush does not merge for you.
FAQ
Questions people actually ask
Can Rush run GPU work like Modal?
No. Rush runs coding and research agents in CPU microVMs. GPU inference, training and RL rollouts are Modal's job.
Can I move from Modal Sandboxes to Rush?
Only Modal's coding-agent use case would move to Rush, so there is no general migration guide. RL rollouts, evals and GPU work stay on Modal.
Does Rush need my own model account?
By default a run uses your own Claude or ChatGPT subscription or API key, and your provider bills the model calls. Rush meters the sandbox compute against your plan, and parked time is free.
Cluster
Keep reading
Citations
Sources
Every factual claim on this page traces to one of these URLs. Retrieval date is 17 October 2026 unless the source prints an earlier published date.
- [1] Rush, “Rush product requirements”, retrieved 2026-09-15. https://getrush.ai/cloud
- [2] Rush, “Rush pricing (live contract on getrush.ai)”, retrieved 2026-09-15. https://getrush.ai/pricing
- [3] Rush, “Isolation, plainly”, retrieved 2026-10-17. https://getrush.ai/docs/concepts/isolation
- [4] Rush, “Results and proof”, retrieved 2026-10-17. https://getrush.ai/docs/concepts/results-and-proof
- [5] Rush, “Usage and billing”, retrieved 2026-10-17. https://getrush.ai/docs/concepts/usage-and-billing
- [6] Rush, “Rush MCP server”, retrieved 2026-10-17. https://getrush.ai/docs/concepts/mcp-server
- [7] Rush, “Automations”, retrieved 2026-10-17. https://getrush.ai/docs/concepts/automations
- [8] Modal, “Pricing: Pay per second for GPUs and CPUs | Modal”, retrieved 2026-10-17. https://modal.com/pricing
- [9] Modal, “Sandboxes (Modal Docs)”, retrieved 2026-10-17. https://modal.com/docs/guide/sandboxes
- [10] Modal, “Sandboxes: Isolated environments for reinforcement learning and AI agents | Modal”, retrieved 2026-10-17. https://modal.com/products/sandboxes
- [11] Modal, “Networking and security | Modal Docs”, retrieved 2026-10-17. https://modal.com/docs/guide/sandbox-networking
- [12] Modal, “Top AI Code Sandbox Products in 2025”, published 2025-07-17. https://modal.com/blog/top-code-agent-sandbox-products
Next
Run the next Task on Rush
Connect a repository. Pick Claude Code, Codex, or OpenCode. Review the PR.