agental

Compare · agental and Modal

agental or Modal? Here is when to pick each.

Both give an agent a sandbox to run code in. Modal's is a gVisor container unless you ask for a VM, and ends within 24 hours; ours is always a VM and runs until you stop it.

In short

Pick Modal for GPUs, a memory snapshot to start new sandboxes from, or sandboxes beside the functions you already run there. Pick agental for runs past 24 hours, a whole VM every time and a much smaller bill, and for an agent that needs SSH, mail or a browser you drive.

Pick Modal if…

  • GPUsModal rents GPUs by the second, from a T4 to a B300, in sandboxes too. We have none.
  • Sandboxes beside the rest of your workFunctions, model inference, volumes and sandboxes on one platform and one bill. We do sandboxes only.
  • A memory snapshot to start new sandboxes fromModal can snapshot a sandbox with its programs still running, as an experiment, kept for 7 days, and start new sandboxes from it. Ours keep memory across a stop of the same sandbox only.
  • Regions to choose fromModal runs in the US, the EU, Asia-Pacific and more, at 1.15 to 1.75 times the price. We run in one region today.
  • Free credit every monthStarter has no fee and $30 of compute each month. Ours has no fee either, and gives $10 of credit once, for a saved card.
  • Very large disks512 GiB by default, up to 3 TiB. Ours go up to 160 GB.

Pick agental if…

  • About a ninth of the priceThe same 2 vCPU and 4 GiB costs $0.027 an hour here and $0.2380 at Modal's sandbox rates.
  • No limit on how long it runsA Modal sandbox ends after 5 minutes unless told otherwise, and after 24 hours at most. Here it runs until you stop it.
  • A VM every time, not gVisorModal's sandboxes run under gVisor unless you choose its VM runtime. Every one of ours has its own kernel.
  • Stop it, and start the same oneThe same sandbox, its disk as it was, and its memory too if you ask. At Modal you snapshot it and make a new one from the snapshot.
  • SSH, from any terminal or editorConnect the way you connect to any server. Modal gives a shell through its own CLI.
  • Its own email addressEach sandbox receives mail. E2B, Daytona, Modal and Vercel don't offer it.
  • Your Playwright drives its ChromeConnect your own browser automation to the Chrome inside the sandbox, from your computer.

By what you are building.

Kinds of agent, and which of the two we would choose for each, Modal included.

Coding agents that work for hoursRuns are never cut at 24 hours, and a month of five agents costs $27.50 rather than $179.35.→ agental
Browser agentsYour own Playwright drives the Chrome inside the sandbox, from your computer.→ agental
A machine per customer, always onThe same sandbox all month, at about a ninth of the price.→ agental
Agents that need a GPUModal rents GPUs by the second, inside sandboxes too. We have none.→ Modal
Teams already running on ModalFunctions, inference and sandboxes on one platform and one bill.→ Modal

The same month at each.

Five coding agents on 2 vCPU and 4 GiB, 8 hours a working day, their disks kept in between.

agental$27.50
Modal$179.35

agental: Free, no fee; $27.50 of top-ups that add $27.50; 880 h at $0.027, 2,770 h stopped at $0.00135. Modal: Starter, $30 credit; $0.1419 per core (2 vCPU) and $0.024 per GiB, an hour.

Side by side.

Every difference we know of, including the ones we lose.

WhatagentalModalBetter at it
Price of 2 vCPU · 4 GiB, an hourEach one's list price, for sandboxes.$0.027$0.2380agental
A month of five coding agents8 hours a working day, plan fees and credit in.$27.50$179.35agental
Longest a sandbox runsBefore it has to start again.No limit5 min by default · 24 h at mostagental
A virtual machine of its ownIts own kernel, not a shared one.Every sandboxgVisor, or a VM if chosenagental
Stopped and started againWhat comes back.The same sandbox, its diskA new one, from a snapshotagental
Memory kept for laterPrograms still running when it comes back.✓ across a stop, 10% of its price✓ as an experiment, 7 daysEven
Fork a sandboxA copy with the same disk.Every sandboxFrom a snapshotEven
The largest diskWithout a sales contract.160 GB512 GiB, up to 3 TiBModal
Your own outbound rulesWhat a sandbox may reach.Every plan✓ by address or domainEven
SSH from any clientYour terminal, your editor, scp.✓A shell through its CLIagental
An email address of its ownMail in to the sandbox.✓—agental
Its Chrome driven from your computerYour own Playwright connects.✓—agental
A desktop for computer useA screen, clicks and keys.✓Build it yourselfagental
Webhooks as sandboxes changeTold when one stops, starts or fails.✓—agental
GPUsBy the second or hour.—✓ T4 to B300Modal
Metrics to your own collectorOpenTelemetry, where you keep the rest.✓✓Even
A create that is safe to retrySent twice, makes one sandbox.✓By a unique nameEven
RegionsWhere sandboxes run.OneUS, EU, Asia-Pacific and more, at 1.15–1.75×Modal
SDKsOfficial libraries.Python, TypeScript, CLIPython, JavaScript, Go, CLIModal

Modal read from modal.com, modal.com/pricing and modal.com/docs/guide/sandbox on 27 October 2026. Tell us if something changed: [email protected].

The same calls, by their names here.

What you use at Modal today, and what it is called in agental's TypeScript SDK. The Python SDK has the same names.

At ModalbecomesAt agental
modal.Sandbox.create(app=app, image=image)agental.sandboxes.create({ template })Answers when the machine is running.
sb.exec(cmd)sandbox.exec(cmd)A command that fails is an answer: read its exit code.
sb.filesystem.read_text() / write_text()sandbox.files.read() / write()Or copy files with scp over SSH.
sb.snapshot_filesystem(), then a new Sandboxsandbox.stop() / start()The same sandbox back, its disk as it was, and its memory with stop({ keepMemory: true }).
encrypted_ports=[port], sb.tunnels()sandbox.expose(port)An HTTPS address for a port, asked for after the sandbox is made.

What we don't do yet.

If one of these decides it for you, Modal is the better choice today.

  • GPUsNot offered.
  • New sandboxes from a memory snapshotA snapshot holds the disk: a sandbox made from it starts its programs again.
  • A second regionNot yet. We will say so here when there is one.
  • Functions and model inferenceWe do sandboxes only.

Questions people ask.

Is agental a drop-in replacement for Modal Sandboxes?

No. The idea is the same (make a sandbox, run commands, read and write files), but the SDK is different, so the calls are rewritten. Most of them map one to one: the table above shows how.

Why is agental so much cheaper?

We run on dedicated servers we rent by the month, not on a cloud's machines rented by the hour. Modal's sandbox rates are also three times its own function rates.

Is an agental sandbox a container?

No. Every sandbox is a virtual machine with its own kernel, so root inside it, Docker included, stays inside it. At Modal that is the VM runtime, which you choose when you make the sandbox.

Does a stopped agental sandbox keep its memory?

When you ask: `sandbox.stop({ keepMemory: true })` keeps its memory as well as its disk, and start puts back every process as it was, in seconds. Writing the memory takes up to about 2 minutes for a large sandbox that is full. Stopped so, it costs 10% of its running price. The memory is not kept across an upgrade of the server it is on: the sandbox then starts from its disk. Modal's experimental memory snapshot is kept for 7 days and starts new sandboxes; ours brings back the same one.

How long can an agental sandbox run?

Until you stop it. You can give it a time limit, and you or the agent in it can move that limit as often as the work needs.

Where do Modal's numbers come from?

From modal.com, its pricing page and its sandbox guides, read on 27 October 2026. If something there has changed, tell us at [email protected] and we will fix this page.

Try it on your own agent, free for a month.

Start building →