Phoenix.new
Phoenix.new is a fully online AI coding agent built by Fly.io for Elixir and Phoenix developers. You describe an app, and the agent builds it inside a disposable remote virtual machine equipped with a real headless browser.
What is Phoenix.new?
Phoenix.new is a batteries-included, fully online AI coding agent tailored to Elixir and the Phoenix framework. It was announced publicly on 20 June 2025 by Chris McCord, the creator of Phoenix and a Fly.io employee, after starting life as a December 2024 weekend project testing whether a coding agent could be written in Elixir. The stated goal is to make LLM agents work as well in Elixir as they already do in Python and JavaScript. Everything runs in the browser: a VSCode-style interface, a shell button, and a live preview of the application being built. Each session gets its own ephemeral Fly Machine with a full root shell, so the agent installs whatever it needs, including APT system packages, without touching your own machine, while Fly.io's infrastructure guardrails apply: hardware isolation, WireGuard, private networks. The main technical differentiator is that the agent is handed a complete headless Chrome browser rather than a screenshot pipeline. Because it is a full browser, the agent sees real page content and JavaScript state, with or without a human present. It also runs the app, watches the logs, runs mix test, spots errors and goes back to work. Applications live in the cloud from the moment they are created: bound ports are detected automatically, previews are served on a phx.run URL, and port forwarding is built in, which removes the usual deployment step before sharing work. GitHub integration is native, with the gh CLI installed by default, authorisation for internal repositories and pull request creation. Phoenix strengths such as clustering, PubSub and presence are put forward for multiplayer games and collaborative applications, and the agent can explore an existing database through $DATABASE_URL or psql to model Ecto schemas, installing a MySQL client on demand. It is not restricted to Phoenix: Rails, Expo React Native, Svelte, Go and other stacks are already installed, and only the system prompt is tuned for Phoenix. A public demonstration generated a Phoenix LiveView Tetris from a first prompt on stage at ElixirConfEU, and the author uses the agent to triage phoenix-core issues. The homepage keeps expectations modest, saying the tool is good at getting basic apps working. Access is only through a Fly.io account, and the site has no other public page.
What it does
- Describe an application in plain language and watch it take shape
- Get a running development server with a shareable preview URL
- Let the agent test its own front-end in a real browser
- Clone a GitHub repository, browse issues and have a pull request opened
- Explore an existing database and derive Ecto schemas from it
- Install APT system packages and mix dependencies with no risk to your own machine
When to use Phoenix.new / When not to
A quick filter to help you decide if Phoenix.new is the right fit.
When to use Phoenix.new
- Elixir and Phoenix developers who want an agent loop as mature as the ones already available in Python or JavaScript.
- Teams prototyping real-time Phoenix LiveView applications and wanting to see them running within minutes.
- Builders of collaborative or multiplayer products who rely on Phoenix clustering, PubSub and presence.
- Developers who need to share work in progress immediately, through a private preview URL on phx.run.
- Maintainers who want an agent working asynchronously, triaging GitHub issues and opening pull requests with or without a human present.
- Developers who prefer a zero-install, browser-only environment over configuring a local toolchain.
When not to use Phoenix.new
- Teams working mainly in other frameworks: the system prompt is tuned for Phoenix today, even though other languages are pre-installed.
- Engineers maintaining large existing codebases: the homepage itself only claims the agent is good at getting basic apps working.
- Buyers who need a public price before committing, since no rate is published and pricing only appears after signing in with a Fly.io account.
- Users looking for a mobile app or an editor extension: the product is browser-only.
- Developers who want to drive the agent programmatically, as no public API is documented for it.
- Anyone unwilling to create a Fly.io account, and anyone under 13, the minimum age set by the Fly.io privacy policy.
How to use Phoenix.new
A typical end-to-end flow, from setup to results.
- Open phoenix.new and sign in with a Fly.io account, the only button on the page.
- Describe the application you want in plain language.
- The agent starts an ephemeral Fly Machine and works inside a VSCode-style interface.
- Follow the build in the live preview, and open a root shell yourself if you need one.
- Let the agent test its own work through the headless browser while it watches the logs and runs mix test.
- Share the generated phx.run URL to show the work in progress.
- Connect a GitHub repository to clone code, read issues or have a pull request opened.
- Provide a $DATABASE_URL so the agent can explore an existing database and propose Ecto schemas.
Pros & Cons
Pros
- The only coding agent built specifically for Elixir and Phoenix, written by the creator of the framework himself.
- A full headless browser instead of screenshots: the agent sees real page content and JavaScript state, which makes front-end verification far more reliable.
- Root shell in a disposable VM: the agent can install anything it needs, including system packages, with no risk to your local machine.
- The application is online from the first minute with a shareable URL, which removes the deployment step from the feedback loop.
- Closed loop of code, logs, tests and correction, with GitHub integration going all the way to the pull request.
- Nothing to install locally, and versatile despite the specialisation, since other languages and stacks are already present.
- Backed by Fly.io infrastructure (hardware isolation, SOC 2 Type 2, ISO 27001 data centres) with a transparent, named public sub-processor list.
Cons
- No public pricing at all: no pricing page, no amount, and no documented free plan or free trial.
- The public site is a single page, with no documentation, no FAQ and no feature page, so there is very little to evaluate beforehand.
- A Fly.io account is mandatory to see anything of the product.
- Deliberately modest positioning on the homepage, which claims the agent is good at getting basic apps working.
- Phoenix specialisation: other frameworks are installed but do not benefit from the tuned system prompt.
- No public API to drive the agent, no mobile application and no browser extension.
- No published commitment on AI model training from customer data, no documented training opt-out, US-only hosting with no EU residency option, and no Article 27 GDPR representative.
Pricing & Plans
No price is published for Phoenix.new. The tool has no pricing page, states no amount, and no archived version of the site has ever carried one; access goes through a Fly.io account and the product's pricing is only visible after authentication. As a result, neither a free plan nor a free trial can be confirmed. What is documented concerns the Fly.io platform rather than the agent: resources such as Micro VMs and persistent storage are billed on a pay-as-you-go basis, with monthly billing in advance under the terms of service. Paid support from USD 29 per month, a USD 99 per month compliance offering (pre-signed BAA plus SOC 2 report) and a 40% discount on reserved machines are Fly.io platform items with no documented link to Phoenix.new. Subscriptions renew automatically unless cancelled (terms of service, section 9.2), and overdue amounts carry 1.5% interest per month. Enterprise enquiries go to sales@fly.io.
- the product exposes no rate card
- no tier and no named plan.
- usage-based billing for resources
- an optional paid support plan
- an optional compliance offering
- and machine reservations.
Data, GDPR & hosting
A consolidated view of how Phoenix.new handles your data.
GDPR overview
GDPR is addressed through concrete instruments rather than a blanket claim. Fly.io publishes a Data Processing Agreement pre-signed by the vendor, explicitly intended for customers subject to the GDPR and active as soon as the customer signs it. It states compliance with the EU-U.S. Data Privacy Framework, its UK Extension and the Swiss-U.S. framework, and has signed standard contractual clauses with sub-processors and affiliates that access data. A public sub-processor list names each provider, its country and its activity. Access, correction and deletion rights are exercised through the account profile or support@fly.io, and privacy complaints receive an answer within 45 days at most. Limits: a Privacy Counsel is named, but no DPO and no Article 27 EU representative, and the legal pages never use the term GDPR, which appears only on the compliance page.
Who owns the data?
The Fly.io terms of service state that Customer Data hosted as part of the services, and all worldwide intellectual property rights in it, is the exclusive property of the customer. The customer grants Fly.io a non-exclusive, worldwide, royalty-free licence to use that data in order to provide and improve the services, plus a limited licence on its trademarks and logos for the same purpose; all rights not expressly granted remain reserved to the customer. Usage and performance data, by contrast, belongs to Fly.io and may be exploited in aggregated, de-identified form. Fly.io is not obliged to back up customer data, backups being the customer's responsibility, and the customer may export its data at any time. The privacy policy adds that employees do not access applications except for security, maintenance, or support with the owner's consent.
Reuse rights
Because the customer owns its Customer Data and can export it at any time, it may reuse and redistribute its own applications and content without asking Fly.io for permission. On the collection side, Fly.io states that it collects information only with consent, keeps it to the minimum needed for the interaction, and does not sell it to third parties. Visitors generate cookies and server logs (browser, language, referrer, timestamp, IP address); account holders provide a username, password and email address, everything else being optional. Sensitive personal data is not intentionally collected, and no advertising is hosted on the site. Analytics run through Google Analytics and PostHog Cloud EU, on logged-out pages only, with no session recording and no capture of clicks, keystrokes or form fields; prior consent is required in the EEA, the United Kingdom and Switzerland, and can be withdrawn through the Cookie preferences control. Fraud and abuse prevention at sign-up collects device and browser signals that cannot be disabled, on a legitimate interest basis. Sharing is limited to sub-processors bound by equivalent restrictions, among them Anthropic for AI-assisted log analysis, support responses and signup classification, and OpenAI for incident summary generation. No published document states whether customer content is used to train AI models.
Data retention & training
Hosting summary
The Fly.io privacy policy states that the information collected is stored and processed in the United States, on servers located there. One exception is documented: marketing analytics data is stored in the European Union by an EU-based processor, PostHog Cloud EU, listed in Germany on the public sub-processor register, while the support provider Plain is listed in the United Kingdom. The overwhelming majority of the published sub-processors are US-based. Phoenix.new itself runs on Fly.io machines: the domain resolves to 169.155.49.177, an anycast address announced by AS40509 Fly.io, Inc. The compliance page states that the data centres used are ISO 27001 certified. International transfers are framed by standard contractual clauses and by Fly.io's certification under the EU-U.S. Data Privacy Framework, its UK Extension and the Swiss-U.S. framework. No EU data residency option is documented for the product, so an organisation that must keep application data inside the European Union will not find that guarantee in the published documents.
Where Phoenix.new works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Phoenix.new.
- Budget is impossible to estimate before signing up: no price is published for the agent, and platform billing is usage-based, so the cost follows machine consumption.
- Subscriptions renew automatically unless you explicitly opt out (terms of service, section 9.2), and overdue amounts carry 1.5% interest per month.
- Fly.io is not obliged to back up your data: backups are your responsibility, and session machines are ephemeral, so push your code to GitHub rather than leaving it on the machine.
- Usage data belongs to Fly.io and can be exploited in aggregated, de-identified form, while nothing is published about whether application content feeds AI model training, and no opt-out is documented.
- Data is hosted in the United States with no EU residency option, which may not suit organisations with strict localisation requirements.
- You are giving root access on a machine to an autonomous agent: read the generated code, dependencies and installed packages before any serious use.
- Account deletion removes the full profile within 30 days once requested, but inactive accounts are never purged automatically, so data can be kept indefinitely if you never ask.
Setup & Integrations
Technical difficulty
Getting started is easy: nothing is installed locally, everything happens in the browser, and the only prerequisite is a Fly.io account used to sign in from the homepage. The agent handles its own environment, packages and development server; as the announcement puts it, you just open the VSCode interface and push the shell button. The real difficulty sits downstream: Elixir and Phoenix knowledge helps when reviewing and correcting the generated code, and advanced uses such as pointing the agent at an existing database through $DATABASE_URL or authorising internal GitHub repositories call for genuine technical skills.
Deployment
Integrations
Supported languages
Behind Phoenix.new
Fundraising
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What is Phoenix.new?
Do I need to install anything, or create an account?
Can I use it for stacks other than Phoenix?
How does the agent check its own work?
Can I share an application while it is being built?
Does it integrate with GitHub?
How much does Phoenix.new cost?
Where is my data hosted?
Is a Data Processing Agreement available, and are sub-processors disclosed?
Is my data used to train AI models?
Should you pick Phoenix.new?
Phoenix.new is a genuinely singular proposition: an AI coding agent built for the Elixir and Phoenix niche, by the creator of the framework himself. The technical choices are strong and coherent, with a disposable root VM the agent can reshape at will, a real headless browser instead of screenshot guesswork, and an application that is online and shareable from the first minute. For a Phoenix developer who wants an agent loop comparable to what Python and JavaScript users already have, there is little else like it. The reservations are just as clear. Pricing opacity is the first obstacle to any evaluation: no rate has ever been published for the agent, and the pay-as-you-go billing documented by Fly.io covers platform resources, not this product. The public surface is minimal, one page with no documentation, no FAQ and no feature page, so you have to create a Fly.io account to judge on the merits. The vendor's own positioning stays modest, claiming the agent is good at getting basic apps working, which points away from large existing codebases. On the trust side, being backed by an established cloud provider helps: a SOC 2 Type 2 report, ISO 27001 data centres, a pre-signed DPA and a named public sub-processor list are more than many young AI tools offer. Undocumented areas remain, and they matter: nothing is published about whether application content feeds model training, no training opt-out is described, and there is no data residency option outside the United States. Worth the detour for Elixir and Phoenix teams comfortable with usage-based cloud billing and US hosting, provided they treat the budget as an unknown until they sign in, and review the generated code before anything serious depends on it.
- Choosing a selection results in a full page refresh.
- Opens in a new window.