NEW: API keys for CI & e2e testing

Isolated sandboxes
for AI agents

Give autonomous agents a real, disposable environment — isolated microVMs, a live terminal, and open models from Hugging Face when you don't have a GPU. Free plan included.

dashboard — openhost.sh
LIVE — sandbox running
$ python3 agent.py --task "scrape pricing page"
→ sandbox oh-9f21ac started, expires in 4m52s
→ writing report.json …
Active agents
3
Plan
Paid
ISOLATED MICROVMSHUGGING FACE MODELSSUPABASE AUTHSTRIPE BILLINGWORKSPACE RLSSSE STREAMINGBYOK VAULT ISOLATED MICROVMSHUGGING FACE MODELSSUPABASE AUTHSTRIPE BILLINGWORKSPACE RLSSSE STREAMINGBYOK VAULT

Running agent code shouldn't be scary

Untrusted, AI-generated shell commands need a real boundary — not a shared container, not localStorage, not a global API key.

Without openhost.sh

  • ✕Agent shell commands run on your own infra, unsandboxed
  • ✕One shared API key, no per-workspace revocation
  • ✕No hard runtime cap — a stuck agent runs forever
  • ✕No audit trail of who launched what, or when

With openhost.sh

  • ✓Every agent gets its own isolated microVM sandbox
  • ✓Workspace-scoped, role-gated, revocable API keys
  • ✓Hard timeout ceiling enforced server-side, every launch
  • ✓Append-only audit log for every sandbox and credential event

Built for agents that actually do things

Isolated microVMs
Every launch boots a fresh, isolated microVM sandbox — never a shared container. If an agent goes rogue, nothing else on the platform is touched.
Live output streaming
Watch stdout/stderr the moment it happens over Server-Sent Events, scoped to exactly the workspace that launched it. The activity feed resumes where it left off after a dropped connection; a live command's raw output does not replay what happened while you were disconnected.
Encrypted credential vault
Save Anthropic or Hugging Face keys once, AES-256-GCM encrypted at rest. HF Inference Providers run the weights; this sandbox is just the agent box. Keys never round-trip to the browser.
Real workspaces, real roles
Owner/admin/member/viewer roles, backed by Postgres row-level security — not a client-side check you have to trust.
Hard-capped run limits
Free plan sandboxes auto-expire at 5 minutes; paid workspaces default to 20 and can push higher. Enforced server-side, not by trusting the client's request.
Scriptable API access
Paid workspaces can mint role-scoped API keys and drive the whole platform from CI, a script, or your own tooling — no browser required.

How it works

From sign-in to a running sandbox in three steps.

1
Sign in & pick a workspace
Email/password or Google, backed by Supabase Auth. Your first workspace is provisioned automatically.
2
Launch an agent
Pick a Hugging Face or Claude template, attach a saved token, and boot an isolated microVM. You don't need local hardware to host the model.
3
Watch it run, live
Stream stdout/stderr in real time, collect artifacts when it's done, and let the sandbox auto-expire.

Built for whoever's shipping the agent

SOLO BUILDER

Ship an agent side project

Free plan gives you one sandbox at a time with a 5-minute cap — pick a Hub model and go. No GPU, no local weights, no credit card.

TEAM

Give your team real workspaces

Invite teammates with owner/admin/member/viewer roles, share a credential vault, and see every launch in one audit-logged workspace.

CI / PLATFORM ENGINEER

Automate it entirely

Mint a workspace-scoped API key and drive sandbox creation from a CI job or your own e2e test harness — no dashboard required.

Why builders choose openhost

The 5-minute free-plan cap forces the sandbox to reap itself automatically — no forgotten VMs racking up cost, no surprise bill on a plan we tell you is free.
Workspace-scoped, role-scoped API keys mean your e2e suite gets its own credential — never a shared master key floating around your team's secrets.
Saved provider API keys are resolved server-side and handed straight into the sandbox's environment — the browser that launched the agent never sees the plaintext key at all.

Start free. Upgrade when you outgrow it.

Two plans, enforced the same way everywhere: server-side, never trusted from the client.

Free
One sandbox, short runs, everything you need to try it.
1 concurrent sandbox
✓5-minute sandbox runtime
✓Live output streaming
✓Encrypted credential vault
✕Programmatic API keys
Start for free
Paid
MOST POPULAR
High concurrency and longer runs for real workloads.
20m+ runtime, customizable
✓Up to 25 concurrent sandboxes
✓20-minute default, customizable runtime
✓Workspace-scoped API keys for CI
✓Everything on Free
Upgrade to Paid

Stop managing infrastructure for your agents

Sign in, launch a sandbox, watch it run. Free plan needs no credit card.

Create your first sandbox