Pi hosting on the AI plan you already pay for
Run Pi, the minimal open-source agent harness, on a persistent, always-on Agent 37 Cloud instance from $4.76/mo. It boots ready on the managed model with no account of your own, and it signs in with a Claude, ChatGPT or Copilot subscription whenever you want one: no other harness here takes all three. Provision it with one API call, hand it long jobs, and come back to a machine that still has the repo, the caches, and the context.
Add a card, nothing is charged, and get a $5 starter credit: first instance free.An agent that teaches itself new tools should not forget them overnight
Pi starts as four tools and grows: you add extensions, skills, and prompt templates until it fits the way you work. On a disposable sandbox all of that is gone by morning, along with the repo, the caches, and yesterday's thread. Hosting Pi on a persistent instance means the environment is part of the work, not a cost you pay again every day.
The usual objection to persistence is price, because the box never turns off. Here that objection is $4.76 a month, metered per minute from a prepaid balance.
A persistent computer behind one API
Agent 37 Cloud provisions isolated, always-on instances from Docker images. Pi runs from the one-click agent37-pi template, ready on the managed model, and the API treats it like any other workload.
One-click template
The agent37-pi template is how Pi runs here: one POST and the instance comes up with Pi preinstalled and ready on the managed model. Custom images remain for advanced setups.
Three ways to pay for the model
Use the managed model and pay per minute, set a provider key and pay that provider, or sign in to a Claude, ChatGPT or Copilot plan you already have. Codex takes a ChatGPT account and Claude Code an Anthropic one; Pi is the only harness here that takes all three, and the only one that needs none of them.
One POST to provision
POST /v1/instances provisions a persistent computer running your image on always-on compute. Each instance is an isolated machine with its own disk.
A URL per instance
Every instance is reachable at https://{id}.agent37.app, with any port routable. No reverse proxy, no DNS, no certificates.
Streaming built in
POST /v1/responses on the instance URL runs a turn and streams text, reasoning, and tool activity over Server-Sent Events.
State that stays
Files, memory, and sessions persist until you delete the instance. Extensions, skills, checkouts, and half-finished branches are waiting next session.
What always-on actually costs
Estimated monthly cost of keeping a roughly 4 GB box on 24/7. Sandbox platforms are great at what they are for; a machine that never turns off is not what they are for.
The 2 vCPU / 4 GB shape is $4.76/mo, and disk expands at $0.09 per GB per month. Everything is metered per minute from a prepaid balance: no seats, no minimums, and billing stops the moment you delete an instance.
No account to connect, no login to wire up
Pi is open source, MIT licensed, and on Agent 37 it runs out of the box on the managed model. Here is everything involved, optional parts included.
The agent37-pi template
Nothing to package: create the instance with the agent37-pi template and it comes up with Pi preinstalled and ready on the managed Agent 37 model. No login, no account to connect. Custom images remain for advanced setups.
An Agent 37 Cloud workspace
Sign in, mint an API key, and provision. Add a card (nothing is charged) and your workspace gets a $5 starter credit, roughly a month of the cheapest instance, so the first instance is free.
Your own credential, if you want one
Optional, and Pi takes two kinds. Set a provider key (ANTHROPIC_API_KEY, OPENAI_API_KEY, a Google or OpenRouter key, and more) when you create the instance and you pay that provider directly. Or open the instance terminal, run pi, and use /login to sign in with a Claude, ChatGPT or Copilot subscription you already pay for. Skip both and the managed model just works.
Schedule it. Ship it under your own brand.
Scheduled and cron runs
Give an instance a cron schedule and it runs unattended: overnight batch fixes, dependency updates on a timer, a code-health report before standup. A persistent machine means every scheduled run starts warm.
White-label by default
Reselling a coding-agent service? You hold the API key, provision one instance per customer, and cap per-instance spend. Your customers see your brand, never Agent 37.
Pi hosting questions, answered
Is there a one-click Pi template?
Yes, available now: the template is named agent37-pi. POST /v1/instances with that template and the instance comes up with Pi preinstalled and ready on the managed model, no login required. Optionally set a provider key in the instance env, or sign in to a Claude, ChatGPT or Copilot plan from the instance terminal, then chat through the API. The persistence, streaming, scheduling, and isolation described on this page all apply.
Can I run Pi on the Claude or ChatGPT subscription I already pay for?
Yes. Open the instance terminal in your browser, run pi, and use /login to sign in with a Claude, ChatGPT or Copilot plan. Codex accepts a ChatGPT account and Claude Code an Anthropic one, but Pi is the only harness here that takes all three, and the only one that also works with no account at all. The credential stays on your instance disk and survives restarts, and we never read it. With a plan covering the model, the machine is the only thing you pay us for: $4.76/mo for a box that never turns off.
What does Pi hosting cost?
$4.76/mo for a 2 vCPU / 4 GB instance awake around the clock, and less when it sleeps: an idle instance pays for its disk alone. Extra disk is $0.09 per GB per month. Everything is metered per minute from a prepaid balance with no seats, and billing stops the moment you delete an instance. Add a card, nothing is charged, and you get a $5 starter credit, so the first instance is free to try. The managed model needs no account of your own, and its usage is metered per token from the same prepaid balance, on top of the instance. Bring your own key or subscription and the model side bills there instead.
Can Pi run 24/7 here?
Yes. Instances are persistent and always-on by design. A refactor that takes all night keeps running after you close your laptop, and cron-style scheduled runs execute unattended so the results are waiting in the morning.
Where does my code and data live?
On your instance's persistent disk, inside an isolated container running under the gVisor runtime with a hard per-instance disk quota. Files stay on the instance until you delete it, and you can delete anytime. If you set a provider key or sign in to a subscription, that credential lives on your instance and goes from there to that provider; we never need it.
Does Pi come with app integrations?
Not yet, and this is the one template where the answer is no. Pi has no MCP client, so the managed Gmail, Slack, and Notion connections that ship with the other harnesses have nothing to plug into. Upstream is working on it, and it will arrive here when it lands. If your agent needs those apps today, host Hermes or OpenClaw instead: same API, same pricing, integrations included.
Is Pi an Agent 37 product?
No. Pi is an open-source agent harness, MIT licensed, built by Earendil Works: a four-tool core (read, bash, edit, write) that extends itself at runtime with TypeScript extensions, skills, and prompt templates. Agent 37 Cloud is the infrastructure layer: a persistent, isolated, always-on computer with an API for provisioning, streaming, files, and schedules. The same platform also hosts OpenCode, OpenAI's Codex, Anthropic's Claude Code, xAI's Grok, the open-source Hermes agent and OpenClaw, and any public Docker image.
Want a different harness on the same machine? The same API runs OpenCode, OpenAI's Codex, Anthropic's Claude Code, and xAI's Grok on the same pricing, plus any public Docker image. Full API details live on the Cloud page.