Formerly KrillinAI, VideoTranslation and Dubbing Tool for 120+ Languages. Open-source AI workspace for creators, powered by Codex. Video downloader, video translation, create videos, images, audios, and edits with Agents in one place. For your Social Media like YouTube, TikTok, Instagram, etc.
OpenRouter model availability
OpenCreator follows OpenRouter's account catalog and explicit access responses. It does not maintain a country allowlist based on where our adapter tests ran. Anthropic text models remain excluded because OpenCreator's current text adapter does not support them. This compatibility constraint is independent of regional access and still applies when OpenRouter includes them in the account catalog.
Official regional policy
OpenRouter's terms, sections 5.7 and 6.8, describe restrictions imposed by individual model providers and the country of the originating request. They do not publish a universal model-by-country table. The provider routing guide links to individual provider terms. A model can also have multiple providers, so its vendor prefix is not a complete access policy.
GET /api/v1/models/user
filters account provider preferences, privacy settings, and guardrails. The
request includes output_modalities=all to cover media as well as text.
This API does not promise to filter every restriction on the user's country.
Filtering through an EU or US regional hostname concerns data residency, not a
general list of supported user countries.
Display policy
- A model must appear in both the gateway's compatible catalog and the actual customer's OpenRouter catalog. Expired or unavailable account metadata hides the models until refreshed.
- Text models must have Responses support in the gateway's compatible catalog
and pass OpenCreator's compatibility policy, which excludes Anthropic models,
OpenRouter routing models, and batch variants. Cached catalogs and default
model selection apply the same policy.
Login, catalog refresh, model selection, chat startup, and Creator preflight
validate the catalog and configuration without additional inference. Only
user-confirmed connection tests send separate paid requests, as described below.
Startup inference also requires explicit test opt-in using
OPENCREATOR_ENABLE_PAID_STARTUP_PROBE=1together withOPENCREATOR_REQUIRE_CODEX_PROBE=1; normal Desktop startup does not enable it. - Image, video, speech, and transcription models are admitted without internal
verificationmetadata. The customer's directory must include the model and its declared output modality must match. Published voices, image parameters from/images/models, and video parameters from/videos/modelsupdate the gateway's capability metadata. Optional discovery failures retain the existing gateway parameters. Internal adapter test records are release evidence, not a customer model allowlist. - Ordinary transcription does not require subtitle timestamps. A subtitle
request using
verbose_jsonmust return actual timestamps; missing timings are rejected without inventing them or resubmitting a paid request. Speech voices are validated against the published catalog regardless of test flags. Speech synthesis and previews use the upstream's default audio encoding; OpenCreator consumes the actual MIME and wraps PCM in WAV using the returned sample-rate and channel metadata. This avoids forcing MP3 on PCM-only models. - An explicit model permission, region, or provider terms failure hides the exact model. After one hour, a successful account catalog refresh can restore it. Quota, moderation, unsupported request formats, rate limits, and temporary outages are not evidence of a geographic restriction. Accepted jobs and error bodies inside successful HTTP responses must also be checked for restrictions.
- Changing the account, key material or version, proxy, or observed network exit invalidates cached catalogs, access failures, and recent success observations. Account/key identities and the network exit are hashed; the customer's key and raw IP address are not persisted in this cache. Exit detection is optional and has a short timeout; failures do not block catalog loading or use of a valid catalog.
Refresh and recovery
Catalogs are cached for five minutes and updated atomically after every page is validated. Login fetches a fresh catalog with the customer's key through the same network transport as generation. Concurrent queries share the same request. Temporary refresh failures can reuse an unexpired catalog for the same account, key version, and network. Expired catalogs hide models until refreshed. HTTP 401 and 403 catalog responses invalidate even an unexpired cached catalog. HTTP 401 also requests a binding refresh through the existing account service to recover a replaced model key, with a 30-second retry limit and no inference.
The account menu's Refresh models command and Reconnect force a catalog refresh. Automatic retries after errors are limited to once per 30 seconds. Logout and account changes cancel in-flight queries, and late results cannot restore the previous account's models. Cache version 3 retains only unexpired exact-model rejections from version 2; old allowed and unknown probe states are discarded.
An unavailable directory is reported separately from a successful empty or incompatible directory. Both leave account sign-in and billing accessible. Automatic Runtime recovery waits until idle. Explicit model selection is allowed while tasks, model requests, and remote jobs are in progress. It updates the default without replacing the same account's Runtime hosts or direct connection: existing turns and Creator job snapshots keep their submitted models, and new tasks use the updated default. Request snapshots and paid job receipts are preserved, with no automatic resubmission after a restriction.
Selection and actual request observations
Official service settings, the chat composer, and Creator tasks accept models that appear in the current account catalog and pass adapter compatibility checks. Selecting a model saves the choice immediately without a probe. Chat startup and Creator preflight validate the selected model and configuration; the task's own inference request is the first paid request.
Catalog membership is not evidence of successful inference. Account snapshots
expose separate model observations: unverified, recent_success, and
restricted. A completed task request records success for its specific
operation, such as Responses or Chat Completions, for five minutes. A success
on one text interface does not establish support for the other. Reading or
replaying a cached result does not refresh its success timestamp.
Explicit model permission, region, or provider terms rejections override parallel successes and hide only that model. Timeouts, insufficient credits, quota limits, rate limits, and temporary service failures clear recent success evidence without quarantining the model. Late results from an old account, key, network context, or changed model metadata cannot update current evidence. Actual requests still handle live provider errors and keep accepted media job IDs and receipts; a connection problem does not trigger another submission.
Explicit connection tests
Each official model row has a connection-test action. It requires confirmation that a real request may consume credits, including when a test fails. Text tests let the user choose Conversation (Responses) or Text processing (Chat Completions), and call only that interface. Cancelling sends no inference. Testing a model does not change the default selection.
Every explicit test uses the current account's inference key and calls
https://openrouter.ai/api/v1 through the existing direct adapter. There is no
probe credential request to the account gateway and no separate platform key.
Valid cached credentials allow a check even when the account gateway is offline
or the account access token has expired, just as they allow normal inference.
Model access, proxy routing, and key permissions therefore match generation.
These tests perform real inference and may consume customer credits; their
result reports costBearer: customer. Failure does not guarantee zero cost,
especially if a remote job was accepted. Probe receipts remain separate from task receipts so they
cannot be replayed as task outputs. HTTP 401 reports expired model credentials,
HTTP 402 and key limits report insufficient credits, and uncertain failures
do not automatically resubmit inference. A confirmed HTTP 401 may refresh the
binding and retry once only when the actual model key has been replaced.
The earlier implementation depended on
POST /api/v1/client/probe-credentials. On 2026-10-14 that endpoint returned
HTTP 404, stopping the check before OpenRouter was called. That dependency has
been removed; probe availability no longer requires a gateway rollout.
- A Conversation test checks a completed Responses tool call. A Text processing test checks a nonempty, completed Chat Completions reply. Each test is one call.
- Image checks one generated, decodable raster using published parameters.
- Speech checks a short delivery with audio MIME and nonempty content.
- Transcription checks a small known spoken WAV and nonempty transcription.
- Video uses minimum published duration and size, waits for completion, and retrieves nonempty video content. Acceptance alone does not pass.
Concurrent identical connection tests share one in-flight request. A new user-confirmed test makes a fresh request. Recent success observations are informational and are never a paid prerequisite for selection or task startup.
These checks test a specific small request, not every capability or parameter. They use the same model key as subsequent tasks. Future requests remain subject to live provider routing, permissions, moderation, and availability; failures must still be handled without automatic paid resubmission.
Live check on 2026-10-14
The same existing customer key returned HTTP 200 and 664 models from the Japan
proxy exit and the US exit reported by the production server's OpenRouter trace.
The sorted ID lists were identical. Both queries used output_modalities=all and
submitted no inference requests. This verifies the catalog integration for these
two exits, not worldwide geographic access.
The initial policy showed 309 text models and hid 135 media candidates because the saved gateway catalog lacked internal adapter verification metadata. The subsequent repair removes that dependency: media discovery uses OpenRouter's catalog and capability APIs and does not require a gateway verification rollout.
The running development account and settings page subsequently showed 309 text, 49 image, 26 video, 35 speech, and 25 transcription models, with all five model selectors enabled. Live platform-funded native-encoding checks passed for Kokoro 82M (24 kHz mono WAV) and HeyGen Voice 1 Turbo (44.1 kHz mono WAV). Both outputs decoded completely and contained non-silent audio. Voxtral Mini Transcribe recognized the known test sentence and returned real segment timings. Both independent test keys were disabled afterwards; customer credits were not used. Management-key accounting confirmed USD 0.00040728 for four speech calls and one transcription call, including the first two speech deliveries before the verification script's decoder path was corrected. Image and video submission, parameter validation, receipts and replay were covered by targeted automated tests, without new paid image or video generations.