What is Open Agents?
Open Agents is an open-source reference app from Vercel Labs for running coding agents in the cloud instead of on your laptop. It has three layers: a Next.js web app for sign-in, sessions and a streaming chat UI; an agent that runs as a durable multi-step workflow on Vercel using the Workflow SDK; and a Vercel Sandbox VM that holds the cloned repo, shell, git and dev servers. The agent sits outside the sandbox and works on it through file, search and shell tools, so the VM can hibernate and resume from a snapshot without ending the run.
A run can clone a repo, work on a branch, and optionally commit, push and open a pull request through a GitHub App. Sessions can be shared as read-only links, sandboxes expose ports 3000, 5173, 4321 and 8000 for previews, and voice input uses ElevenLabs. Models are called through Vercel's AI Gateway. A deploy button provisions Neon Postgres; you add a Vercel OAuth app and a GitHub App yourself.
The repo is MIT-licensed with about 5,900 GitHub stars. It has no tagged releases; the last commit to main was on 4 June 2026, and the most recent pull request activity was on 29 August 2026.
Is Open Agents worth it?
Open Agents is worth reading for one design decision: the agent is not the sandbox. Most background-agent setups run the model loop inside the VM, so a dead VM is a dead run. Here the agent is a durable workflow that survives across requests and talks to a plain VM through tools, so the sandbox can sleep between turns and come back from a snapshot. Anyone building their own Codex-style cloud agent gets a working example of streaming, cancellation, GitHub App auth and auto-PRs in one repo.
It is a template, and it says so. Setup means a fork, a Vercel project, Postgres, a generated auth secret, a Vercel OAuth app and a GitHub App with three callback URLs, and it only runs on Vercel's own Workflow, Sandbox and AI Gateway. Against OpenHands, which you can self-host anywhere and which drives several agent backends, it is narrower and more opinionated.
The catch: cost and upkeep. Vercel's Hobby plan includes 5 hours of Sandbox Active CPU a month and caps sessions at 45 minutes; Pro bills $0.128 per Active CPU hour and $0.0212 per GB-hour of memory once the $20 credit is spent. An issue filed in June reports a solo deployment using about 12 hours of Active CPU over 5 days with nobody using it, blaming sandboxes that never hibernated, eager provisioning and a 30-minute idle window, and suggests the hobby resource profile. Main has had no commit since 4 June 2026, so fixes like that are still waiting, with recent work sitting in unmerged branches.
Our take: treat it the way you would a good conference talk with source attached. Fork it to learn how durable agents and sandboxes fit together, set spend alerts before the first deploy, and look elsewhere if you want a background agent that someone keeps patched for you.
Where it earns its place
A platform team at a 40-engineer company already on Vercel Pro
Offer an internal agent that turns issues into pull requests without anyone's laptop running
Fork the repo, connect a GitHub App installed on the org, and let the agent open PRs for review. Sandbox time on Pro is billed at $0.128 per Active CPU hour after the $20 credit; set Spend Management alerts first and consider the hobby resource profile for small teams.
A developer building an agent product
Learn how to make long agent runs survive timeouts and redeploys
Read packages/agent and the workflow code in apps/web to see how a turn spans persisted workflow steps and how the sandbox resumes from a snapshot. No deployment needed; pnpm web runs it locally with a filled .env.
Before you commit
- Setup
- Fork, import into Vercel with the deploy button (provisions Neon Postgres), set BETTER_AUTH_SECRET, deploy, create a Vercel OAuth app, then a GitHub App with homepage, callback and setup URLs, and redeploy.
- Integrations
- Vercel Workflow SDK, Vercel Sandbox, Vercel AI Gateway, Better Auth with Vercel and GitHub sign-in, GitHub App for repos and PRs, Neon Postgres, optional Redis or Vercel KV, optional ElevenLabs voice input.
- Data and privacy
- Self-hosted in your own Vercel account: sessions in your Postgres, code in your sandboxes, prompts to the models behind AI Gateway. Shared session links are read-only. The web fetch tool requires approval since a May 2026 change.
- Cost at scale
- On Vercel Pro, sandbox usage first draws on the $20 monthly credit; beyond it, Sandbox Active CPU is $0.128/hour and provisioned memory $0.0212/GB-hour in iad1, plus AI Gateway model tokens. Hobby includes 5 Active CPU hours and 420 GB-hours, then pauses sandbox creation.
- Limits
- Hobby sandboxes stop after 45 minutes per session and allow at most 4 vCPUs; Pro allows 24 hours and 8 vCPUs. No tagged releases; 57 open issues and pull requests at the time of checking.
What’s good
- Agent runs outside the VM, so sandboxes hibernate without killing runs
- Full stack in one repo: UI, workflows, sandbox, GitHub App, auto-PR
- MIT licence; written to be forked and changed
- Read-only session sharing and preview ports built in
What’s not
- Tied to Vercel Workflow, Sandbox and AI Gateway
- Long setup: Postgres, OAuth app, GitHub App and secrets
- Reported idle sandbox usage on default settings
- No releases; no commit to main since June 2026
Best for
Engineers on Vercel who want a working reference to build their own background coding agent from.
Open Agents alternatives
Open Agents: common questions
- Is Open Agents worth it?
- A clear, well-argued blueprint for cloud coding agents on Vercel, best read as code to fork rather than a product to run, and one whose sandboxes need watching on the bill. We rate it 3.6 out of 5.
- How much does Open Agents cost?
- Free template (MIT). Running it bills Vercel usage: on Pro, Sandbox Active CPU $0.128/hour and memory $0.0212/GB-hour after the $20 monthly credit, plus model tokens via AI Gateway.
- What are the downsides of Open Agents?
- Tied to Vercel Workflow, Sandbox and AI Gateway; long setup: Postgres, OAuth app, GitHub App and secrets; reported idle sandbox usage on default settings; no releases; no commit to main since June 2026.
- Who is Open Agents best for?
- Engineers on Vercel who want a working reference to build their own background coding agent from.
- How much does Open Agents cost for a team or in production?
- On Vercel Pro, sandbox usage first draws on the $20 monthly credit; beyond it, Sandbox Active CPU is $0.128/hour and provisioned memory $0.0212/GB-hour in iad1, plus AI Gateway model tokens. Hobby includes 5 Active CPU hours and 420 GB-hours, then pauses sandbox creation.
- Is Open Agents safe for company data?
- Self-hosted in your own Vercel account: sessions in your Postgres, code in your sandboxes, prompts to the models behind AI Gateway. Shared session links are read-only. The web fetch tool requires approval since a May 2026 change.
- What are the best Open Agents alternatives?
- The closest alternatives we have reviewed are OpenHands, OpenCode, OpenAI Codex CLI and Vibe Kanban.
Answers checked Nov 20, 2026.