Siray.ai is an OpenAI-compatible API gateway that routes one key to 698 catalogued endpoints from nineteen upstream vendors, spanning language, image and video models. Its current positioning centres on uncensored variants, and it publishes neither terms nor a privacy policy.
Used this tool? Rate it
Used this tool? Rate it
Siray.ai is a unified AI API gateway that puts many third-party generative models behind a single key and a single billing relationship. Rather than opening accounts with OpenAI, Anthropic, Alibaba, ByteDance and a dozen others separately, you point your code at one endpoint and select a model with a a provider-slash-model identifier such as black-forest-labs/flux-kontext-i2i-pro. The company markets this as 300+ models through one API; counting the public documentation index directly yields 698 catalogue entries across nineteen named upstream providers, so the marketing figure counts distinct models conservatively while the catalogue counts individual endpoints.
Two things about this product need stating plainly at the top, because they shape everything else on this page.
First, the current positioning is not what you might expect from a developer infrastructure service. The home page headline reads Uncensored AI, Private Creation, and the site title is "Siray | Uncensored & Private AI". The pitch is access to models without content filters and without refusals. Roughly a fifth of the catalogue consists of endpoints labelled "Spicy" or "Uncensored". This is a deliberate market position, not an incidental feature, and it is the single most important fact for deciding whether the service fits your project. Note also that the directory description we inherited for this listing describes something noticeably different — a general cost-focused API gateway with discount offers — with no mention of the uncensored positioning at all. This page follows the live site.
Second, the operator is not identified anywhere. No company name appears anywhere on the home page, the documentation, or the contact page; the only copyright symbol on the home page belongs to an image credit for the Rijksmuseum in Amsterdam. The LinkedIn company page lists Singapore, self-employed, and two to ten people. There is no terms of service and no privacy policy at any reachable address. For a metered API that takes payment and processes your prompts and outputs, that is a material gap, and it is treated as such throughout this page.
The catalogue spans OpenAI, Anthropic, Google, Alibaba, ByteDance, Kuaishou, Black Forest Labs, xAI, MiniMax, DeepSeek, Tencent, Midjourney, Vidu, PixVerse, Z-AI, Lightricks, Xiaomi, Tripo and Moonshot AI. Endpoint counts skew toward OpenAI (39), ByteDance (33) and Alibaba (25). As a multi model gateway the practical benefit is straightforward: one credential, one invoice, one integration to maintain instead of a dozen.
The documentation site is titled "OpenAI Compatible API" and the advertised migration path is to point your existing code at a new base URL rather than rewrite it. Authentication uses a bearer token in a SIRAY_API_TOKEN environment variable against https://api.siray.ai/. For teams already built on the common convention, this is the difference between an afternoon and a sprint.
There is a published siray package on npm, MIT licensed, currently at version 1.3.0 and first published on 10 November 2025, plus a matching siray package on PyPI at 0.3.0. Both are real, installable and independently verifiable — a useful sanity check for any small vendor, since a landing page costs nothing while a maintained SDK does not.
Generative image and video calls follow a job pattern: you submit a job, then poll for the result. The Node SDK exposes pollIntervalMs and timeoutMs and returns a status object with isCompleted() and isFailed(). The HTTP path is a POST to an async generations endpoint followed by a GET against the returned task id. This matters for architecture — you need somewhere to hold state between submission and completion, which a synchronous chat integration does not require.
Both SDKs provide a helper that turns a local file into a base64 data URL suitable for the image field, so image-to-image and image-to-video work without you first hosting the source file somewhere public.
The vendor states intelligent routing across providers with roughly 10ms added latency, 99.9% availability, 15+ global data centres on the home page and 30+ acceleration nodes in the documentation. All of these are self-reported with no status page, audit or third-party measurement offered, and two of them contradict each other (see limitations).
Roughly 80 of those catalogue entries carry a Spicy or Uncensored label, offered as variants alongside the standard versions of the same models — Alibaba's Wan and Qwen Image families, ByteDance's Seedance and Seedream families, MiniMax and others. The documentation describes these variants as supporting "broader, less-filtered creative exploration".
Against that positioning, the official blog carries an explicit compliance statement: the platform prohibits all illegal content, holds zero tolerance for CSAM, requires that all generated images depict legal adults aged 18 or over, and defines "uncensored" as meaning no moderation beyond what the law requires — never permission for illegal content. The company also states elsewhere that certain endpoints are moderated on every generation and that no prompt set is guaranteed to pass.
The most defensible use is evaluation. When you do not yet know whether a task suits one vendor's model or another's, paying one bill to try several beats opening accounts everywhere. Model selection is a string change.
A team already calling three or four providers carries three or four contracts, invoices and key-rotation procedures. Collapsing that to one is real operational savings independent of any per-token discount.
If your product already speaks the common chat convention, the same key extends to an image generation API and a video generation API without a second integration. The asynchronous job pattern is the one architectural addition required.
The catalogue includes Chinese-origin model families — Qwen, Wan, Seedance, Seedream, GLM, Hunyuan, Kling — alongside Western ones. For teams outside those vendors' home markets, aggregation can be simpler than arranging access individually.
This is the use the site itself foregrounds, and it should be stated rather than skirted. Teams building adult-oriented creative products are the audience for the Spicy and Uncensored lines. Anyone doing this takes on the legal obligations directly: the platform's stated floor is the law plus its CSAM and 18+ rules, and everything above that floor is your responsibility, in a service that publishes no terms of service defining the relationship.
For high-volume generation where per-call price dominates, an aggregator claiming to route to the cheapest viable path is worth benchmarking — though you will have to benchmark it yourself, since no public price list exists.
Registration is on the main site; API keys are managed at the console. The documentation compresses onboarding to two steps: create an account, then manage keys in settings.
npm install siray or the PyPI equivalent, or skip both and call the REST endpoints directly. All three paths are documented on the home page with working snippets.
The documentation explicitly warns against hardcoding the secret and directs you to pass SIRAY_API_TOKEN through the environment.
Identifiers take the vendor-prefixed form. The documentation cautions to always take current identifiers from the live catalogue, which implies they do change — worth pinning and monitoring rather than assuming stability.
Submit the job, then poll. Set a sensible poll interval and timeout; the documented example uses three seconds and a three-minute ceiling. Treat a timeout as an expected state, not an error case.
Because there is no public price list and no status page, the vendor's savings and uptime claims cannot be checked in advance. Run your own comparison against direct vendor pricing on your actual workload before migrating anything that matters.
If you touch the Spicy or Uncensored lines, write down what your product will and will not generate, and how you will enforce it. The platform's floor is legal compliance plus its stated 18+ rule; it does not substitute for your own controls.
Verify the price on your own workload first. The savings figures published are 30, 40 and 70 percent in three different places, and the price table itself does not render publicly. Treat all three as marketing until your own metering says otherwise.
Pin model identifiers and watch for changes. The documentation's own advice to always use the latest identifiers is a hint that names move. A pinned identifier plus an alert on failures is cheaper than a silent breakage.
Keep a direct-vendor fallback path. With no published status page or audited uptime, a single aggregator is a single point of failure. Since the surface is OpenAI-compatible, keeping the ability to swap the base URL back to a direct provider costs little and buys a lot.
Do not send regulated or highly sensitive data. With no privacy policy, no data processing agreement and no named legal entity, there is nothing to hold anyone to. That is a sufficient reason to keep protected health information, personal data of EU residents and confidential client material out of it.
Ask support for the terms in writing before any commercial commitment. Email and Discord are the available channels. If you intend to build a business on this, the absence of published terms is the first thing to resolve, not the last.
Treat the model documentation as a catalogue, not a specification. Descriptions are visibly template-generated, with identical sentences repeated across many unrelated models. Use them to discover what exists, then verify actual behaviour by calling the endpoint.
Check whether an endpoint is actually live. The company's own blog notes that documentation for some variants appears before those variants reach the live catalogue. Availability in the docs does not guarantee availability in production.
Budget for the async pattern. If your current integration is entirely synchronous, adding image or video generation means adding job state, retry handling and timeout policy. Plan that work rather than discovering it.
Developers and small teams evaluating models get the clearest benefit: broad access at low commitment, useful precisely during the phase when you do not know what you need.
Indie builders and startups watching per-call costs may find the aggregation economics work, provided they measure rather than assume.
Teams needing image and video alongside text avoid a second integration, which is a genuine time saving.
Builders of adult-oriented creative products are the audience the site addresses most directly, and they should read the limitations section closely before committing.
Who should look elsewhere: any organisation with compliance obligations, since there is no privacy policy, no data processing agreement and no identifiable counterparty; anyone who needs contractual uptime guarantees, since the availability figures are self-reported and internally inconsistent; anyone who cannot proceed without knowing prices in advance; and any team whose procurement process requires a named legal entity.
This is API-only infrastructure. There is no desktop application, no mobile app and no browser extension. There is no consumer-facing application here at all.
REST API — the primary interface, at https://api.siray.ai/, authenticated with a bearer token.
Node.js SDK — the siray package on npm, MIT licensed, version 1.3.0. Four releases exist: two on 10 November 2025, one on 19 December 2025 and one on 23 January 2026. The package declares neither a repository nor a homepage.
Python SDK — the siray package on PyPI, version 0.3.0, MIT licensed, described as a Python client library for the Siray AI API.
Web console — at console.siray.ai, for account and key management. It is a login-gated application and its contents cannot be inspected without an account.
Documentation site — at docs.siray.ai, which also publishes an llms.txt index designed for AI agents to enumerate every page.
Support — an email address and a Discord invite. There is no phone number, office address or ticketing system, and the contact page offers only a four-field form.
The main marketing site itself is minimal: its sitemap contains exactly three URLs — home, pricing and contact.
Siray.ai is positioned as pure pay-as-you-go with no contracts and no monthly minimum. Beyond that, honesty requires saying what could not be established.
No usable price list is public. The pricing page presents three tables for chat, image and video endpoints, each with columns for model name, creator, price and actions — and the price table itself renders empty. No individual model documentation page carries a price either; a keyword sweep for currency symbols and cost terms across sampled model pages returned nothing. Rather than repeat figures from third-party pages that could not be verified, this page states the position plainly: check the live pricing page and your console after signing in, because there is no reliable public number to quote.
The savings claims are inconsistent across the site. The home page says 30 percent average savings in one place and up to 70 percent in another; the pricing page says up to 40 percent. Three different figures, all self-reported, none audited.
A free entry point exists but is unquantified. The home page offers "Start for Free" and model pages carry a "Get Free APIs" call to action, yet no public page states how much free usage that actually includes, for how long, or under what limits.
There is no published billing agreement. With no terms of service, questions like refund policy, price-change notice, minimum charges and account termination have no published answer. Anyone spending meaningfully here should get those in writing from support first.
Going direct to the model vendors — OpenAI, Anthropic, Google and the rest — costs you the aggregation convenience but gives you published pricing, published terms, status pages, data processing agreements and a named counterparty. For anything regulated or commercially significant, that is a materially different risk profile.
Established aggregators occupy exactly this niche with longer track records and public documentation of their commercial terms. Anyone attracted by the multi-model proposition should compare on published pricing and published policies, not only on catalogue size.
Self-hosting open-weight models suits teams with the infrastructure skills and steady volume to justify it. You trade convenience for full control of both cost and content policy — the latter being the actual reason many teams look at uncensored endpoints in the first place.
Other unrestricted-API services exist; this is a small but real category rather than a single offering. If unfiltered access is your requirement, compare providers on what they publish about legal compliance, corporate identity and data handling, since those are exactly the dimensions on which this category varies most.
No identifiable operator. No company name appears anywhere on the site, documentation or contact page. The LinkedIn entry gives Singapore, "Self-Employed" as company type, and two to ten employees — which sits oddly beside documentation promising forward-deployed engineers and enterprise-grade compliance. You cannot know who you are contracting with.
No terms of service and no privacy policy. Four candidate paths return HTTP 404 and the sitemap contains no legal pages. The site does claim SOC 2 compliance, zero-data-retention redundancy and "full-chain compliance" in marketing copy, but none of it links to a policy, an audit report or a data processing agreement. Marketing claims of compliance without published documents are not verifiable.
Uncensored endpoints are named after the upstream vendors whose models they wrap. An endpoint titled "Wan 2.7 i2v Uncensored by Alibaba Cloud" offers less-filtered output from a model whose original vendor generally restricts exactly that. The site publishes no licensing explanation and no non-affiliation statement. It does disclose its upstream suppliers openly and in detail, which is more than many such services do, but users should understand that the relationship with those vendors is undocumented.
Self-reported reliability figures contradict each other. The home page states 99.9 percent in two places and 99.99 percent in a third, and the documentation returns to 99.9. Similarly, the home page claims 15+ data centres while the documentation claims 30+ acceleration nodes. No status page exists to check any of it.
No pricing transparency. Covered above, and worth restating as a limitation: you cannot compare this service against alternatives before creating an account.
Documentation is template-generated. Many model descriptions repeat word-for-word across unrelated models. The pages are template-generated rather than individually written, so they establish that an endpoint exists but tell you little about how it actually behaves.
Some documented endpoints are not live. The company's own blog states that developer documentation for certain variants appears ahead of those variants reaching the live catalogue. Verify availability before designing around a specific endpoint.
Effectively no third-party track record. Trustpilot shows zero reviews and no score at all, on an unclaimed profile. Searches of the major technology outlets returned no independent review or reported article. The npm timeline explains why: the first SDK shipped on 10 November 2025, under ten months before this page was written, with four releases and the most recent more than seven months old at the time of writing. This is an early product, and this page deliberately excluded affiliate review blogs rather than pad the evidence.
Same-name confusion is likely in searches. Several unrelated companies share the Siray name — a Quebec staffing agency founded in 1992, a dissolved UK restaurant company, an active UK retail company registered in December 2022, and an Italian design agency. None has any connection to this product. Every fact on this page is anchored to the siray.ai domain and its verified official properties.
Content responsibility sits with you. The platform states a legal floor and a CSAM and 18+ rule. Everything else — jurisdictional rules on synthetic adult content, likeness and consent, platform distribution policies — is your problem, and there is no published contract allocating any of it.
One API key reaching a large multi-vendor catalogue. The company advertises 300+ models; counting the public documentation index gives 698 entries across nineteen named providers including OpenAI, Anthropic, Google, Alibaba, ByteDance, Kuaishou, Black Forest Labs, xAI, MiniMax, DeepSeek, Tencent and Midjourney. Coverage spans language models, text-to-image, image-to-image, text-to-video, image-to-video and reference-to-video.
The documentation site is titled "OpenAI Compatible API" and the advertised path is to change your base URL and key rather than rewrite your integration. Model identifiers use a vendor-prefixed form. For chat-style calls the migration is genuinely small; for image and video you additionally need the asynchronous job pattern, which is a real if modest piece of work.
This cannot be answered from public information. The pricing page's model tables render empty, and no model documentation page lists a price. The site claims savings of 30, 40 and 70 percent in three different places. There is a free entry point, but its size and limits are not published. Check the pricing page and your console after registering, and benchmark on your own workload.
They are variants offered alongside standard versions of the same models — roughly 80 of the 698 catalogue entries — described as supporting broader, less-filtered creative content. The uncensored positioning is the site's headline proposition rather than a hidden corner of the product. If your project has no need for that, the same platform still serves the standard endpoints, but you should know what you are signing up to.
Yes, and it is explicit where it exists. The official blog states that the platform prohibits all illegal content, holds zero tolerance for CSAM, requires all generated images to depict legal adults aged 18 or over, and defines "uncensored" as no moderation beyond what the law requires, never permission for illegal content. The company also states that some endpoints are moderated on every generation regardless. What is missing is a formal terms of service in which any of this is contractually binding.
Publicly, this cannot be determined. No company name, registration or address appears on the site, documentation or contact page. LinkedIn lists Singapore, self-employed, and two to ten employees. Crunchbase was blocked to first-hand verification and its content was therefore not used. For procurement processes that require a named counterparty, this is likely to be disqualifying on its own.
There is no way to establish that from published information. The site markets private creation, zero-data-retention redundancy and SOC 2 compliance, but publishes no privacy policy, no data processing agreement and no audit report to support any of it. Absent those documents, treat the service as unsuitable for regulated, confidential or personal data.
Unknown from public evidence. The claimed figures are 99.9 percent in the home page metrics and documentation and 99.99 percent in another home page section — an internal contradiction — with no status page, incident history or third-party monitoring. Given a single-vendor dependency and no service commitment, keeping a direct-provider fallback configured is prudent.
Real, yes; actively maintained is less clear. The npm package is MIT licensed at version 1.3.0 with four releases: two on 10 November 2025, one on 19 December 2025 and one on 23 January 2026. The PyPI package sits at 0.3.0. Both install normally. The most recent release was over seven months before this page was written, and the npm package declares neither repository nor homepage, so there is no public issue tracker to gauge responsiveness.
That depends entirely on what the system does. For prototyping, model comparison and non-sensitive workloads it is a reasonable convenience. For anything carrying compliance obligations, handling personal or confidential data, or requiring contractual guarantees, the absence of a named operator, published terms, privacy policy and verifiable pricing are serious obstacles — not because the service is necessarily bad, but because you would be unable to demonstrate diligence to anyone who asked. At minimum, get the commercial terms in writing from support before you depend on it.
