Niches · Code execution sandboxes

Run a coding agent harness on a repository in a cloud box

Two or three companies' docs show a full fit.

2companies, full fit
16companies, fit with a limit
24companies carry the tag
21share of the category's work
11first-hand accounts

What counts as done

The harness starts inside the sandbox with git and the language toolchain available, the task runs to a commit or a diff that is returned or pushed, the session survives the agent thinking for 15 minutes without being stopped, and output streams to the caller.

How agents ask for it: run Claude Code on this repo in the cloud, let Codex work on this branch somewhere safe, ship this fix from my phone, run the tests the agent wrote

Does the job 2

The docs show it meets the test above, and nothing we read says otherwise.

ListingWhy
Agent 37 Cloud C
Agent 37 Inc.
Templates ship Hermes, Claude Code, Codex, OpenCode and other harnesses with git; auto_sleep is off by default so nothing stops a long run; exec runs in the background and turns stream over POST /v1/responses. A launch-thread user found setup easy. source
Freestyle B
Freestyle Cloud Inc.
Ubuntu base snapshots ship Node, Bun, Python, Docker, git and coding agents on PATH; there is no idle timeout unless set, exec runs up to 24 hours on updated guests, and PTY sessions stream output. source

Does the job, with a limit 16

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
Blaxel Sandboxes C
Blaxel, Inc.
Claude Code and Codex tutorials, Dockerfile images and volumes exist, but a sandbox with no open connection goes to standby in about 15 seconds, freezing a thinking harness unless it runs with keepAlive and timeout 0. Broccoli runs one-shot coding jobs on it. source
Celesto (formerly SmolVM) BB
Celesto AI, Inc.
Persistent computers with files and state across sessions run on the owner's machine or Celesto Cloud, and nothing stops a long run, but no harness image, git toolchain or output streaming is documented. source
Cloudflare Sandbox SDK BB
Cloudflare, Inc.
The coding-agent runner guide clones a repo, runs Claude Code in the background and returns a diff, but code inside does not count as activity: the instance stops at the inactivity timeout unless an alarm keeps it, and it bills while running. source
Daytona BB
Daytona Platforms Inc.
Snapshots build from any Docker image and volumes persist, but the 15-minute auto-stop default fires even with processes running; set autoStopInterval 0. One team built a janitor and session store around crashes (HN 47320329); issue 5154 reports buffered activity ignored. source
Deno Sandbox D
Deno Land Inc.
A user runs Claude Code in yolo mode inside a snapshot-based sandbox, but the maximum lifetime is 30 minutes, there is no custom image, and the volume and snapshot CLI had sharp edges (missing --from). source
E2B BB
FoundryLabs, Inc.
A Claude Code template installs the CLI on a Node image with git, but Hobby sandboxes stop after 1 hour of continuous running (24 hours on Pro at $150 a month), and the 5-minute default timeout kills unless set or on_timeout is pause. Workboxes runs OpenCode sessions on it. source
InstaVM D
Checksum Labs, Inc.
Snapshots from any OCI image, volumes and SSH exist, but no lifetime, idle or streaming behaviour was documented on the pages read, and the Free plan runs 5 VMs. source
Maritime B
Maritime AI, Inc.
Templates for Claude Code, Codex, Hermes or any Docker image, /data kept, exec detached up to 3,600 seconds. Idle sleep counts API activity, not guest work, so set --idle high or the harness is snapshotted mid-run. source
Modal Sandboxes B
Modal Labs, Inc.
Any registry image and volumes, but the 5-minute default and 24-hour maximum lifetime mean snapshotting and restarting for long work, as Restate's durable agent does; a codex exec in an A100 sandbox hit the orphan bug. source
Morph Cloud D
Morph Labs, Inc.
Pause keeps memory and ttl_action pause avoids losing work, but no harness image, idle setting for plain instances or output streaming was documented, and compute needs a paid plan. source
SPACE C
Perplexity AI
Templates such as seed:latest, suspend after six idle hours by default rather than kill, and ECW accepts the E2B Python SDK, but no harness template or git toolchain is documented and only Python is tested. source
Runloop Devboxes B
Runloop AI, Inc.
Blueprints from a Dockerfile, an OpenCode tutorial, gateways for Claude Code and Codex, and no idle action by default, but trial keep-alive is 1 hour and the service ends. source
Sandbox0 C
Lougao, LLC
Digest-pinned OCI images and a coding-agents guide with Codex, Claude Code and OpenCode, but a ttl pause is filesystem-only and discards processes; memory pause is experimental. source
Sprites B
Fly.io, Inc.
Ubuntu 25.10 ships Claude CLI, Codex, Cursor and git, but a Sprite pauses when nothing is using it, so a thinking agent needs a Tasks API hold with a heartbeat. One user left because Claude Code needed a login for every feature. source
Together Code Sandbox C
Together Computer, Inc.
A Dockerfile snapshot carries git and a toolchain and a sandbox runs until terminated, but every port is public without the experimental policy, and access is by request to Together. source
Vercel Sandbox BB
Vercel Inc.
The universal image ships Node, Python and coding agents and persistence keeps the filesystem, but Hobby sessions end at 45 minutes (24 hours on Pro). VoicePatch runs the Claude Agent SDK in it. source

Doesn't do this job 4

Listings in the category that do something else, with the reason.

ListingWhy
Amazon Bedrock AgentCore Code Interpreter BB
Amazon Web Services, Inc.
Pre-built Python, Node and Deno runtimes only, no custom image, and session files are removed at the end; a harness with git and a toolchain is not what the service has. source
DeltaBox Lite C
DeltaBox (delta-box)
A local runtime for a developer's machine; there is no cloud box, hosted session or streaming service. source
Microsoft Execution Containers BB
Microsoft Corporation
An SDK that embeds in an application on the user's machine; no cloud box, session or streaming service. source
Offscreen B
Mark Phelps (individual)
A per-job macOS VM for UI tests on the developer's Mac, not a cloud box for a harness. source

What stops agents finishing

  • a 15-minute idle default stops the sandbox while the model is still thinking (Daytona)
  • session caps on free tiers: 45 minutes on Vercel Hobby, 1 hour on E2B Hobby, 30 minutes on Deno
  • the harness needs a browser login inside the box every time (Sprites account)
  • the harness crashes in streaming mode and the caller must build session storage and cleanup (Daytona account)

Around it

Before: agent-harnesses.start-session, search.find-url. After: agent-harnesses.open-pull-request, code-sandboxes.dev-server-preview.

Where this comes from

Weight 21 of 100 in Code execution sandboxes: 11 of 53 mentions, the joint largest group. Nine Hacker News accounts and one GitHub issue describe running Claude Code, Codex, OpenCode or a Claude Agent SDK loop inside Daytona, Modal, Sprites, Vercel, E2B or Agent37.

The same as data: /api/v1/use-cases/code-sandboxes.coding-agent-workspace.json, or the MCP tool get_use_case with id code-sandboxes.coding-agent-workspace. Other jobs in this category: /api/v1/niches?category=code-sandboxes.

Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

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.