Questions and answers about Ephemeral

Ephemeral is an API, CLI and MCP server that runs an AI agent task in the background inside a fresh, sealed microVM with only the files, API connections, web access and budget you allow, streams every step, and destroys the machine when the task is done. You start runs from the REST API, the CLI, the MCP server, the dashboard or a cron schedule.

Ephemeral is for developers and teams who add background AI agent work to their products and operations. Typical tasks are investigating a failed payment for support, watching a page or a metric, writing a scheduled report, processing data, and running code you do not trust.

Ephemeral is a hosted agent service: you describe a whole task, and Ephemeral runs the agent in a sandboxed microVM and returns the result. It is not a framework you embed in your own code, and not a bare sandbox that your own agent loop drives step by step. A self-hosted edition runs the same service in your own cloud account.

Yes. Ephemeral is a remote MCP server at https://app.ephemer.al/mcp, so Claude, ChatGPT, Cursor and other MCP clients can start runs, wait for results and manage workflows and schedules, signing in with OAuth or an API key. A run can also call your own MCP servers as connections, limited to the tools you allow.

A run can use models from several AI model providers, at the provider's price plus 15%. You choose the model for each run or workflow, and the dashboard and GET /v1/models list every available model with its prices.

The website, documentation and dashboard are available in seven languages: English, Japanese, German, Spanish, French, Brazilian Portuguese and Korean. A few pages, such as the comparisons, use cases, glossary and changelog, are in English only.

A run costs a $0.005 run fee once its worker starts, plus $0.0002 for each second of run time, its model calls at the provider's price plus 15%, and $0.001 per web fetch or $0.015 per web search. There is no subscription or minimum, and a run never costs more than its budget. Every run's page shows its exact cost.

No, the hosted service has no free trial and no free credit; you start by buying prepaid credits from $5. The self-hosted Team license has a 14-day free trial, one per organization.

A run cannot go over its budget. Ephemeral reserves the whole budget before the run starts, refuses any model call or web request that could take the run past it, and releases what the run did not use. The default budget is $0.50, and you can set up to $100 for a run.

Each web fetch, a page or API request, costs $0.001, and each search costs $0.015, with no markup. They are charged from the run's budget like model calls, and web access is off unless you turn it on for a run.

Zero data retention adds 10% to the whole run, including the run fee, compute, the marked-up model calls and web requests. The surcharge counts toward the run's budget.

With prepaid credits, bought by card through Stripe from $5 at a time, with no subscription. Credits never expire, are shared by everyone in your organization, and can top up automatically when the balance runs low.

No. Credentials never enter the machine: a broker outside it adds them to each request and lets through only the HTTP methods or MCP tools you allow. Credential values are also removed from responses before the agent or the run's record sees them.

Each run gets its own fresh microVM, which is destroyed when the run ends. The machine has no direct internet access and none of your credentials, and nothing is shared between runs or organizations.

On the hosted service, data is stored in Japan and encrypted in transit and at rest. A run's definition, output and events are kept for 30 days after it ends, then deleted. With zero data retention, the output is returned once and none of the run's content is kept, only metadata such as its status and cost.

Only if you turn on web access for the run, and then only public sites or the domains you allow. Requests are sent by Ephemeral's own bot, EphemeralBot, not by the machine, and each one passes a safety check before it is sent.

A run can last up to one hour (3,600 seconds). If you set no timeout, it stops after five minutes. When the timeout ends, the worker is stopped and its machine destroyed.

Yes, with the self-hosted edition: the same engine and dashboard as one image, installed in your own AWS account under a license. The Team license costs $490 a month or $4,900 a year for 1 installation, 10,000 runs a month and 5 runs at once, with a 14-day free trial. Business costs $5,000 a month for 2 installations, 100,000 runs a month and 20 runs at once, with add-ons.

Code sandboxes give your own agent loop a machine to start and control step by step, while Ephemeral runs the whole agent task for you. You send a prompt, files, connections and a budget; Ephemeral runs the agent in a sealed microVM, keeps credentials and web access outside it, and returns the result. The comparison pages set out the differences in detail.

Yes. The changelog lists the changes to the service, the API and the CLI, and new CLI versions are published as signed releases, which the ephemeral update command installs.