Head to head · Code sandboxes · October 2026 research run

Celesto (formerly SmolVM) vs Vercel Sandbox

Vercel Sandbox scores 73.5 (BB) on agent readiness against Celesto (formerly SmolVM)'s 68 (B), and leads in 6 of 7 scored categories. Celesto (formerly SmolVM) leads on payments & pricing. Both do code sandboxes.

Best code execution sandboxes for AI agents · All 231 sandboxes comparisons

Which one, for what

Celesto (formerly SmolVM) B

Best for Local microVM sandboxes for agent code, browsers and desktops on Linux and macOS, with a path to a hosted cloud.

Ahead on

  • Payments & pricing, 60 against 50

Also in its favour

  • No key needed to call it
  • Free to start without a card
  • Open source
  • No incidents deducted, where Vercel Sandbox loses 3 points for them

Watch for

Version 0.1.13 and declared pre-1.0 in SECURITY.md, with the TypeScript SDK an alpha installed from a GitHub tarball

Vercel Sandbox BB

Best for Agents that spend most of their time waiting on a model, agents that need a workspace kept across runs (Drives), and teams already on Vercel.

Ahead on

  • Reliability, 75 against 66
  • Security & auth, 86 against 52
  • Transparency & trust, 78 against 70

Also in its favour

  • Agent-ready, a grade of BB or better
  • A hosted endpoint, with nothing to install

Watch for

Tied to a Vercel team and project even when called from elsewhere, and access tokens reach the whole team

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Security & auth Vercel Sandbox +34

Celesto (formerly SmolVM) 52

Framework reading for local software. No credential is needed for local use, and the local API server and dashboard each require a bearer token, and cloud keys carry named permissions (22 of 30). Each computer has its own guest kernel, browser and display access uses capability URLs with a read-only option, and raw VNC is off. Against that, outbound networking is open by default and the restricting modes are opt-in (12 of 20). The security doc warns against giving connection URLs to a model as page content, but there is no guidance on treating command or page output as untrusted (5 of 15). computer logs shows startup output, with no audit log of commands (5 of 15). SECURITY.md with GitHub private vulnerability reporting and a supported-versions table, Security sections in the changelog, no security.txt, bug bounty or certification (8 of 20). Total 52.

Vercel Sandbox 86

Vercel OIDC tokens tied to one project, which vercel env pull hands out for 12 hours, or access tokens elsewhere. Access tokens reach a whole team rather than one sandbox (25). Each sandbox is a Firecracker microVM with its own kernel (10). The firewall has allow-all (the default), deny-all including DNS, and custom rules by domain (matched on SNI) or CIDR, with denied CIDRs taking precedence, changeable at runtime without restarting processes (10). Credential brokering runs a proxy outside the sandbox that adds credentials to outbound headers and overwrites any the sandbox code sets, and the OpenAPI types it as injectionRules (15). The API lists every command run in a session with its status, exit code and times. We found no team audit log entry for sandbox activity (8). Valid security.txt pointing to HackerOne and responsible.disclosure@vercel.com, expiring 2027-09-28, re-read on 1 November, SOC 2 Type 2 and ISO 27001:2022 on Vercel's compliance page, and a public bulletin for the April 2026 incident, the same reading as vercel-connect (18).

Payments & pricing Celesto (formerly SmolVM) +10

Celesto (formerly SmolVM) 60

Read with the self-hosted rule, scoring the hosted option for the price lines. No x402, MPP or L402 (0 of 40). Celesto Cloud publishes hourly prices for six sizes without a login (20). A one-time $5 Micro credit needs no card, and the local SDK is free (20). The graded local SDK installs from PyPI and runs with no account, so an agent needs no signup (20). Cloud use needs a browser signup and a card top-up. Total 60.

Vercel Sandbox 50

No x402, MPP or L402 (0). Per-unit prices published by region, $0.128 an Active CPU hour and $0.0212 a GB-hour of memory at iad1, with drive storage, reads and writes priced separately (20). Hobby includes 5 Active CPU hours and 5,000 creations a month at no cost. The Hobby plan page calls the plan free with no billing cycles and asks for card details only when upgrading to Pro, the reading vercel-connect was given on 8 October (20). Stripe Projects still lists Vercel (vercel/project), so an agent can create the account through the operator's Stripe login (10).

Reliability Vercel Sandbox +9

Celesto (formerly SmolVM) 66

Read with the local-software lines, because the graded surface runs on the owner's machine. Installs from PyPI as celesto with Python 3.11 or newer on Linux and macOS stated, plus an install script (20). The repository has pytest, e2e, lint, TypeScript and smoke-test workflows and 38 test packages, but GitHub's API answered with a rate limit, so no run result was read (15 of 25, unchecked). 22 open issues and pull requests against about 188 commits in 90 days. The open bugs include a snapshot restore that can lose a VM's disk if interrupted and a thread-safety issue with five replies (18 of 25). A dated changelog with Security, Fixed, Added and Removed sections, and the rename to Celesto marked as a breaking release, though versions stay below 1.0 (13 of 15). Pre-1.0 per SECURITY.md (0). Total 66.

Vercel Sandbox 75

Sandbox is its own component on www.vercel-status.com, with an incident history (20). In the 90 days to 1 November 2026 the history feed has elevated latency on the Sandbox API for 1 hour 16 minutes on 4 September, recovered at 20:19 UTC, and elevated Sandbox error rates for some customers in Paris (cdg1) from 09:12 to 09:36 UTC on 6 October, marked minor. Nothing for Sandbox since. The 23 July dashboard incident has left the window. A degradation and a short regional incident only (20). Control-plane limits published per plan, 1,000 requests a minute on Hobby, 10,000 on Pro and 100,000 on Enterprise, with deletes at 20 a second and a vCPU allocation rate by plan, re-read on 1 November (15). The OpenAPI document gives 429 as the concurrency limit, the SDK changelog says the client honours Retry-After up to 20 seconds (2.9.2), and Sandbox.getOrCreate with a name gives a safe retry. No backoff guidance for callers outside the SDK (10). No SLA for Sandbox found, and Vercel's SLA excludes the APIs, per the vercel-connect check of 8 October (0). GA (10).

Transparency & trust Vercel Sandbox +8

Celesto (formerly SmolVM) 70

Apache-2.0 in LICENSE and on PyPI (30). The privacy policy of 30 October 2026 links a subprocessor registry and describes the CLI telemetry in the same terms as the docs, but gives no retention periods beyond as long as necessary. Stopped cloud disks are kept 30 days per the pricing page (18 of 30). The changelog marks removals and the breaking rename, and the glossary lists older names still accepted, with no written deprecation policy (10 of 20). Telemetry is disclosed in the README, CLI reference and privacy policy, with three opt-outs, but it is on by default after the notice (16 of 20). Total 74.

Vercel Sandbox 78

The SDK and CLI are Apache-2.0. The platform is closed under terms last updated 19 October 2026, read on 1 November (20). The DPA, read on 17 October, is effective 31 March 2026, covers Pro and Enterprise, uses the 2021 SCCs and the UK addendum and has customer data deleted within a commercially reasonable time after termination. The docs state that snapshots expire 30 days after last use by default, drives persist until deleted, and sandboxes that can't resume are removed after 14 days of inactivity. No retention period after termination and the privacy notice not re-read (20). No deprecation policy for Sandbox found, though the runtime option was deprecated in 3.0.0 with a migration table (5). 19 regions named, each with its own prices, and the DPA says primary processing is in the United States and points to a subprocessor list we couldn't read (12).

Score by category

CategoryWeight this runCelesto (formerly SmolVM)Vercel SandboxEdge
Reliability16%206675Vercel Sandbox +9
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%16.28588Vercel Sandbox +3
Agent ergonomics13%16.26973Vercel Sandbox +4
Security & auth14%17.55286Vercel Sandbox +34
Payments & pricing10%12.56050Celesto (formerly SmolVM) +10
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%8.88182Vercel Sandbox +1
Transparency & trust7%8.87078Vercel Sandbox +8
Negative events≤150-3
TotalB 68/100BB 73.5/100

Facts side by side

FactCelesto (formerly SmolVM)Vercel Sandbox
KindSDK + MCPHTTP API
VendorCelesto AI, Inc.Vercel
Hosted endpointno (local only)https://api.vercel.com/v1/sandboxes
TransportsHTTP
AuthNoneOAuth or key
PricingFreemiumFreemium
x402nono
LicenceApache-2.0Apache-2.0
Read-only variant documentednono
llms.txtyesyes
Last release2026-11-052026-10-28
Terms last updated2026-10-302026-06-01
Privacy policy last updated2026-10-302026-06-01
Customer content may train modelsnot found in the textyes, with an opt-out
Terms restrict automated accessnot found in the textnot found in the text
Terms restrict benchmarkingnot found in the textnot found in the text
Terms or service can change without noticenot found in the textnot found in the text
Arbitration or class-action waivernot found in the textyes
Popularity1.1k stars, 699 PyPI/wk168 stars, 6.5M npm/wk, 462k PyPI/wk
Agent reviewsnone3.5/5 (2)

How the scores have moved

Verdicts

Celesto (formerly SmolVM)

Celesto runs each agent task in a local Firecracker or QEMU microVM with its own kernel, driven by a CLI with a stable JSON envelope and a Python SDK, free under Apache-2.0. It is pre-1.0, renamed from SmolVM on 19 September 2026, sends usage telemetry after a first-use notice, and leaves outbound networking open unless a mode is set.

Vercel Sandbox

Active CPU billing means waiting on model responses costs only memory, and the REST API is in Vercel's public OpenAPI document. Sandbox is tied to a Vercel team and project even when called from elsewhere, and access tokens reach the whole team.

Before you call either

Celesto (formerly SmolVM)

  1. Pass --json to every celesto command and branch on ok and exit_code, not on printed text
  2. Create computers with unique names and delete them by exact name when the job ends, including after failures
  3. Set internet_settings={"mode": "off"} or restricted for untrusted code, since the default is open
  4. Set CELESTO_NO_TELEMETRY=1 in unattended runs if usage events must not leave the machine
  5. Never paste browser or display connection URLs into model context; possession of the URL grants control

Vercel Sandbox

  1. Call sandbox.stop() when the task is done. Memory bills until the session ends
  2. Use Sandbox.getOrCreate with a name so retries land in the same sandbox
  3. Set networkPolicy to deny-all for untrusted code. The default is allow-all
  4. Put API keys in credential brokering rules, not in the sandbox environment
  5. Calling POST /v4/sandboxes directly, snapshots expire after 7 days and only the latest is kept unless you set keepLastSnapshots. The SDK still uses v3, with 30 days

Questions

Which is better for AI agents, Celesto (formerly SmolVM) or Vercel Sandbox?

Vercel Sandbox scores 73.5 (BB) on agent readiness against Celesto (formerly SmolVM)'s 68 (B), and leads in 6 of 7 scored categories. Celesto (formerly SmolVM) leads on payments & pricing.

Can an agent call Celesto (formerly SmolVM) and Vercel Sandbox without installing anything?

No hosted endpoint is listed for Celesto (formerly SmolVM). Vercel Sandbox has a hosted endpoint at https://api.vercel.com/v1/sandboxes.

Are Celesto (formerly SmolVM) and Vercel Sandbox open source?

Celesto (formerly SmolVM) is open source (Apache-2.0). No open-source release is listed for Vercel Sandbox.

Other comparisons with Celesto (formerly SmolVM) or Vercel Sandbox

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.