Proxed.AI
An open-source AI gateway for iOS apps that call OpenAI, Anthropic or Google AI directly. Change a single URL and your provider key stays split between server and app, guarded by Apple DeviceCheck and schema-validated responses.
What is Proxed.AI?
Proxed.AI is a secure AI gateway that sits between an application and the model provider it calls. The problem it addresses is specific: an API key embedded in a mobile binary can be pulled out by decompiling the app or by proxying its traffic, and the site names the consequences plainly, from leaked data to runaway bills and compliance violations.
The fix is deliberately unglamorous. You change one URL. Instead of calling OpenAI, Anthropic or Google AI directly, your app calls the matching Proxed.AI endpoint, and three mechanisms take over. The first is a split-key architecture: the sensitive half of your provider key stays on Proxed.AI's server while the app carries only a harmless fragment, sent in an x-ai-key header. The second is Apple DeviceCheck: the app attaches a cryptographic token in an x-device-token header, Proxed.AI validates it, and emulators, scripts and jailbroken devices are refused before they consume any quota. The third is structured responses: a drag-and-drop Visual Schema Builder defines objects, arrays, enums and required fields, every response is checked against the schema, and invalid ones are regenerated.
Structured endpoints cover text, vision and PDF, with image generation alongside them and pass-through routes for each provider. Supported providers are OpenAI (GPT-4.1, GPT-4o, GPT-4o-mini and the GPT-5 family), Anthropic (Claude 4 Opus and Sonnet) and Google AI (Gemini 2.5, 2.0 and 1.5). The home page cites product recognition, medical imaging, entity extraction, sentiment analysis, classification, invoice parsing and research-paper analysis as use cases.
Underneath, a Hono and Bun data plane handles traffic while a Next.js control plane runs the dashboard. Retries use exponential backoff with jitter on 429, 502, 503 and 504 responses, per-provider circuit breakers are exposed on a health endpoint, SSE streaming is handled explicitly, and latency and retry headers report on each call. Logs, latency and per-run cost are tracked live, and usage notifications fire on a call threshold within a chosen window.
Everything is open source under the nech-ai/proxed repository and can be self-hosted for free with no usage limits. The service is published by Comonad Limited in the United Kingdom, founded and led by Alex Vakhitov; it launched publicly on 18 February 2025 and its changelog runs to 2 January 2026.
What it does
- Route AI calls to OpenAI, Anthropic or Google AI by changing a single URL
- Split the provider key so the sensitive half stays on the server and only a harmless fragment ships in the app
- Verify the hardware authenticity of the calling device through Apple DeviceCheck and reject failures with a 403
- Enforce a response schema and automatically regenerate responses that fail validation
- Turn images, text and PDFs into structured output through dedicated vision, text and PDF endpoints
- Generate images through an experimental endpoint covering gpt-image-1, dall-e-3, dall-e-2 and imagen-4.0
- Track usage, latency and cost per run in real time, and raise an alert above a chosen call threshold
When to use Proxed.AI / When not to
A quick filter to help you decide if Proxed.AI is the right fit.
When to use Proxed.AI
- iOS and Swift developers who currently embed an OpenAI, Anthropic or Google AI key straight into their app binary
- Indie developers and early-stage projects, the audience the Starter plan is explicitly written for
- Product teams already in production that need a hard ceiling on their monthly AI call volume and spend
- Security and platform engineers who want provider secrets out of the client and key rotation handled in one place
- Privacy-focused teams, regulated industries and enterprise deployments that need to self-host the whole stack and read the source
When not to use Proxed.AI
- Non-technical end users: this is developer infrastructure with no consumer-facing interface
- Teams shipping Android-only or purely web front ends, since DeviceCheck hardware attestation exists only on Apple devices
- Anyone counting on a permanent free hosted tier or on very high volumes: there is no free hosted plan and the top tier stops at 50,000 calls per month
- Organisations that require a signed DPA, a named sub-processor list or SOC 2 / ISO 27001 evidence, none of which is published
- Anyone looking for an AI model itself: Proxed.AI supplies none and simply routes requests to OpenAI, Anthropic and Google
How to use Proxed.AI
A typical end-to-end flow, from setup to results.
- Create an account on the Proxed.AI dashboard to start the 14-day trial, which needs no credit card, or deploy the self-hosted edition instead
- Create a project in the dashboard and note the project ID it is given
- Configure your provider key: the server half is stored by Proxed.AI, the client fragment is copied into your app
- Build your response schema in the drag-and-drop Visual Schema Builder, preview sample responses and validate it
- Replace the provider URL in your app with the matching Proxed.AI endpoint, appending your project ID
- On iOS, obtain a DeviceCheck token with DCDevice.current.generateToken() and send it in the x-device-token header
- Send your key fragment in the x-ai-key header, or use the documented Bearer alternative combining the partial key and the device token
- Decode the already-typed response straight into a Decodable struct in your app
- Set a call threshold and a time window in the project settings to receive usage alerts by email
- Follow runs, latency, costs and models used from the dashboard, and keep an eye on the public status page
Pros & Cons
Pros
- Minimal implementation: one URL change, no backend to build, with the site claiming a working setup in under five minutes
- Fully open source and free to self-host, with no usage limits and no vendor lock-in
- The complete provider key never exists inside the shipped app binary
- Apple DeviceCheck hardware attestation delivers a level of protection that is hard to build yourself
- Typed, schema-validated model output with automatic regeneration when a response fails validation
- Three providers behind a single interface, with the cost per call shown on each plan (0.0021, 0.0008 and 0.0005 USD)
- 14-day trial with no credit card and a 30-day money-back guarantee, alongside a detailed changelog, a public status page and open architecture documentation
Cons
- No permanent free hosted tier: the site says so in plain terms
- Attestation is limited to Apple's ecosystem, so Android and web clients get no DeviceCheck protection
- Modest call quotas on the hosted plans, at 1,000, 10,000 and 50,000 calls per month
- No DPA, no named sub-processor list and no SOC 2 or ISO 27001 certification are published, and the word GDPR is never used on the site
- Incomplete documentation: the Quickstart page reads Coming soon, and the rate-limiting page reads Coming Soon while the paid plans already sell rate-limiting rules
- A very small structure: a single-director UK company with no announced external funding
- Adds an intermediary in the critical path of every AI request
Pricing & Plans
There is no permanently free hosted plan; Proxed.AI states explicitly that it does not offer one. Every hosted plan includes a 14-day free trial with no credit card required. The lowest paid entry point is the Starter plan at USD 25.00 per year on annual billing, exclusive of VAT and covered by a 30-day money-back guarantee. The site's metadata separately advertises a starting price of USD 2.50 per month for monthly billing, and the yearly option is presented as 25% cheaper. A free alternative exists outside the subscription: the open-source edition can be self-hosted at no cost and with no usage limits.
- free. The open-source edition deployed on your own infrastructure
- with no usage limits
- no subscription and no vendor lock-in.
- USD 25 per year
- roughly USD 0.0021 per call. One project
- 1
- 000 API calls per month
- DeviceCheck authentication
- real-time monitoring
- a basic analytics dashboard and email support.
- USD 100 per year
- roughly USD 0.0008 per call and advertised as 83% cheaper than Starter. Unlimited projects
- 10
- 000 calls per month
- DeviceCheck
- advanced rate-limiting rules
- real-time logs and monitoring
- custom limit configurations and priority email support.
- USD 300 per year
- roughly USD 0.0005 per call and advertised as 50% cheaper than Pro. Unlimited projects
- 50
- 000 calls per month
- DeviceCheck
- advanced analytics and cost tracking
- detailed usage reports
- custom integrations
- Monthly and yearly billing are both offered
- with the yearly option announced as 25% cheaper. Prices exclude VAT and every plan carries a 30-day money-back guarantee.
- Custom plans are available on request through the contact link on the pricing section.
Data, GDPR & hosting
A consolidated view of how Proxed.AI handles your data.
GDPR overview
The word GDPR never appears anywhere on the site: not in the privacy policy, not in the terms, not on any product page. No compliance claim is made in so many words. The privacy policy nevertheless follows the regulation's structure without naming it. It sets out legal bases (contractual necessity, legitimate interests, consent, legal obligation) and grants access, rectification, erasure, restriction, objection, portability and withdrawal of consent. Transfers outside the United Kingdom or the EEA are covered by Standard Contractual Clauses or an adequacy decision, and users are told they may complain to their local supervisory authority, the UK Information Commissioner's Office being given as the example. No Data Protection Officer is named, no Article 27 EU representative is designated, no DPA is published, and sub-processors are described only by category. Rights requests go to privacy@proxed.ai. The policy took effect on 2 May 2025.
Who owns the data?
The service belongs to its publisher: Comonad Limited (Company Number 15713725) and its licensors keep all rights to the platform's code, design, trademarks and text, under the law of England and Wales. The terms say nothing about who owns the content you push through the proxy or get back, so ownership of customer data is left unaddressed rather than claimed. You remain responsible for the provider keys you configure, and full API keys are not retained in retrievable form. Personal data may be shared with hosting, payment, analytics and email providers, with authorities and in a business transfer, but is never sold or rented.
Reuse rights
Nothing in the terms claims a licence over what you send through the proxy or over what the models return, so outputs appear to be yours to use, publish and commercialise without asking Proxed.AI for permission. That freedom rests on silence rather than on an explicit grant, which is worth noting: the terms neither confirm nor restrict reuse. Proxed.AI says it processes your data to deliver and maintain the service, to analyse usage patterns for product improvement, to communicate with you and to meet legal obligations. Training models on customer data is not mentioned anywhere on the site, in either direction. Because the models belong to OpenAI, Anthropic and Google, their own reuse conditions apply and Proxed.AI does not address them. The product's own source code is open and can be self-hosted with no usage limit.
Data retention & training
Hosting summary
Proxed.AI does not say where its data is hosted. There is no trust or security page, and the privacy policy names no country, no region and no cloud provider location. What is documented is the jurisdiction: the service is run by Comonad Limited, a UK company, and the terms are governed by the law of England and Wales. The privacy policy states that personal data may be transferred to and processed in countries other than your own, and that transfers outside the United Kingdom or the EEA are covered by Standard Contractual Clauses or an adequacy decision. The architecture documentation names the components, with Supabase for persistence, Trigger.dev for background jobs and a Next.js dashboard, but never their regions. A June 2025 post announced multi-region hosting on Fly.io as coming soon and in testing; nothing has confirmed it since. Self-hosting is the only way to control the jurisdiction, since you supply your own Supabase, Vercel and GitHub accounts. One caution on network data: the domain's public IP is a Cloudflare anycast node, a CDN point of presence rather than a storage location.
Things to keep in mind
Risks and trade-offs to weigh before adopting Proxed.AI.
- The proxy sits in the critical path of your requests: if it goes down, your app's AI features stop working
- The sensitive half of your provider key is entrusted to a third party and reassembled server-side on every request
- The publisher is a very small UK company with a single director and no announced funding, so continuity risk deserves an honest assessment
- No DPA, no sub-processor list and no security certification are published, and GDPR compliance is never claimed by name even though request metadata and IP addresses are processed: contractual due diligence remains entirely on you
- DeviceCheck covers neither Android nor the web, so assuming the whole application is protected would be a false sense of security
- Cost guardrails only work if thresholds are configured correctly, and hitting the monthly quota of 1,000, 10,000 or 50,000 calls can interrupt a production service
- The terms allow Proxed.AI to suspend or terminate access at any time, with or without a stated reason, and cap liability at the amount paid over the previous twelve months
Setup & Integrations
Technical difficulty
Developer-only, but light on the hosted plans. You need Swift or plain HTTP: create a project, split the key, build the schema, change the URL and add two headers, x-ai-key and x-device-token, the latter obtained from Apple's DeviceCheck API on a supported Apple device. The site claims a working setup in under five minutes with no SDK and no backend, which is plausible; note however that the documentation's Quickstart page still reads Coming soon. Self-hosting is considerably harder: Supabase, Vercel and GitHub accounts, database migrations to apply, and retry and timeout environment variables to tune.
Deployment
Integrations
Behind Proxed.AI
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What does Proxed.AI actually do?
Do I need to install an SDK or build a backend?
Is my provider key still inside the app?
How are emulators and bots blocked?
Which providers and models are supported?
Does it work for Android or the web?
Is there a free plan?
How much does the hosted service cost?
Can I self-host it?
What data does Proxed.AI keep?
Should you pick Proxed.AI?
Proxed.AI is a narrow piece of infrastructure that does one thing seriously: it stops a mobile app, an iOS app above all, from carrying a provider API key that anyone can extract. Instead of promising vague security, it names three mechanisms: a split key whose sensitive half never leaves the server, Apple DeviceCheck hardware attestation that rejects emulators and scripts before they burn quota, and schema validation that guarantees typed, parsable model output. All three are verifiable, and the integration cost is a single URL change.
The open-source licence and the free, unlimited self-hosting option are the strongest structural argument in its favour: whatever happens to the hosted service, the code and the deployment remain yours, which largely removes the lock-in that usually comes with putting a third party in your request path.
The reservations are real. There is no permanent free hosted tier, and the company says so explicitly. Monthly call quotas are modest, topping out at 50,000. No DPA, no named sub-processor list and no security certification are published, and the word GDPR appears nowhere on the site even though the privacy policy implements its structure. Some pages still read Coming Soon, including a rate-limiting page whose feature the paid plans already sell. And the publisher is very small: a single-director UK company with no announced funding, which matters when the service sits in the critical path of your app.
The project is clearly alive, with a changelog running to 2 January 2026 and a public status page. For an iOS team that wants provider keys out of its binary today, it is a credible answer, provided the self-hosting escape hatch stays part of the plan.
- Choosing a selection results in a full page refresh.
- Opens in a new window.