← All help topics

AI providers & API keys

Where to get each model's key, how to set it up, how the models differ and what each is best for.

1Where things live: Profile and Project

Keys and model choice live in two places:

  • 🔑 In «Profile» you add providers and their keys (grouped by region) plus your own OpenAI-compatible model (endpoint, model names, token) — stored encrypted, once.
  • ⚙️ In the project («Project» → «Models & limits» → «Which model does what») you set which model handles which task — no keys. Easiest: pick a default model and press «Apply to all tasks».

Without a configured model, a local fallback is used.

2✨ The easiest way — OpenRouter in one click

If you'd rather not register with every provider, add a card and copy keys — connect OpenRouter:

  1. In «Profile» → «Model providers» click «Connect OpenRouter».
  2. You'll be redirected to OpenRouter — sign in and authorize.
  3. The API key (with a built-in spending limit) is saved here automatically, no copy-paste.

One key unlocks hundreds of models (OpenAI, Anthropic, Google, open-source) — in the project pick the «OpenRouter (aggregator)» model and optionally set a specific model in the model name (e.g. openai/gpt-4o-mini). Billing and the limit live in your OpenRouter account. OpenAI/Anthropic/Google themselves don't offer a one-click OAuth to mint a key — so for a fast start this is the best path; manual entry of an individual provider's key still works too.

3🔑 A separate key with a spending limit (important)

Rules for handling keys safely:

  • 🔒 In your provider's console, create a SEPARATE API key specifically for TestIP and set a spending/token limit (usage limit) on it.
  • ⛔ Don't use keys without a limit.
  • 🔁 If you suspect a key is compromised, revoke it with the provider immediately and replace it.

The service tries to save tokens (economy mode, weekly token budget) but cannot guarantee the exact amount charged by the provider or 100% storage security. Keys are stored encrypted (AES-256-GCM) and are never shown after saving.

4🤖 AI can make mistakes

Text and images are produced by generative models and may contain inaccuracies, errors or outdated facts. Always review materials before publishing — every post goes out only after your approval, and responsibility for publishing remains with you.

5YandexGPT — text 🇷🇺

Two ways to connect — pick whichever suits you.

① Via OAuth token (simpler, recommended for newcomers): in «Profile» → «Model providers» pick «Yandex OAuth token», click the «get a token» link — on the Yandex ID page sign in and authorize, paste the token into the field. Also set «Yandex Folder ID». TestIP exchanges the OAuth token for a short-lived IAM token and refreshes it automatically — no service account or key needed.

② Static API key: console.yandex.cloud → create a folder and a service account with role ai.languageModels.user → «Create API key»; fill the «YandexGPT API key» and «Yandex Folder ID» fields.

In both cases the Folder ID is on the folder page in console.yandex.cloud (next to its name, format b1g…).

Traits: strong in Russian, cheap lite tier, data stays in Russia. Best for: Russian-language content and 152-FZ compliance.

6GigaChat (Sber) — text 🇷🇺

Connect in three steps:

  1. Key: developers.sber.ru/studio → «GigaChat API» → create a project → «Get key» → copy the Authorization key (Client ID:Client Secret in Base64).
  2. In «Profile» pick «GigaChat» from the list and paste the Authorization key. TestIP exchanges it for an access token via OAuth 2.0 (client credentials) and refreshes the token automatically.
  3. Pick the access scope: «Individuals (PERS)» — for private users (default, free tier), «B2B» — businesses billed by packages, «CORP» — businesses on subscription.

The scope must match the one issued in the Sber portal, otherwise the token exchange returns an error. You can change it later in «Profile» («Save scope» button).

Traits: Russian, enterprise-oriented. Best for: RU content, business communication.

7DeepSeek — text 🇨🇳

🔗 Key: platform.deepseek.com → API keys (OpenAI-compatible).

🧠 Traits: very cheap, strong reasoning, long context.

✅ Best for: analytical and long-form texts, budget savings.

8Qwen (Alibaba) — text 🇨🇳

🔗 Key: Alibaba Cloud Model Studio / DashScope → API key (international dashscope-intl endpoint).

🧠 Traits: strong multilingual support, competitive price.

✅ Best for: multilingual content, Asian markets.

9Anthropic Claude — text 🇺🇸

🔗 Key: console.anthropic.com → API Keys.

🧠 Traits: excellent writing quality, precise tone-of-voice adherence, long context.

✅ Best for: premium copywriting and complex style requirements.

10OpenAI GPT — text 🇺🇸

🔗 Key: platform.openai.com/api-keys.

🧠 Traits: versatility, reliable JSON mode, good price/quality balance.

✅ Best for: general tasks and consistent output.

11Embeddings model (separate) — meaning-level dedup

Embeddings are text vectors; they let the service tell that two materials are “about the same thing” even when worded differently. Meaning-level dedup in “Feed quality” and source relevance scoring rely on them.

⚙️ Where: «Profile» → «AI models & keys» tab → «🧭 Embeddings model». You can set a separate key, base URL and model name (any OpenAI-compatible /embeddings endpoint — OpenAI itself, your own server, etc.). ⚠️ Enter the BASE URL, without a trailing /embeddings — it is appended automatically. And use the provider's OpenAI-compatible address rather than its native API: e.g. for Alibaba DashScope that is https://dashscope.aliyuncs.com/compatible-mode/v1 (model text-embedding-v4), not the /api/v1/services/embeddings/… path from their curl samples. If the compatible mode is unavailable, switch «API format» to «Alibaba DashScope (native)» — then you can paste the address straight from the curl sample. The «Test» button reports which URL and format were used and what the provider actually replied. So one model can write texts and a cheaper one can compute embeddings.

🧩 If you don't set a dedicated model, embeddings come from your OpenAI or Yandex key when present. Anthropic, DeepSeek and GigaChat have no embeddings — then meaning-level dedup is off (the project settings show a warning).

📌 In project settings («Models & limits» → «Feed quality») you can override just the embeddings model name per project; the key and endpoint come from the profile.

12🧭 Checking the embeddings model

Embeddings are how the service understands the MEANING of text. Two things rely on them: matching material to the project topic, and “meaning-level dedup” (not publishing a second post with the same message).

🔎 To check it works: “Profile” → “AI model access” → the “🧭 Embeddings model” block → the “Check” button. It does NOT merely look at whether a key is set — it actually computes a vector and shows the result: which provider answered, the vector's dimensions and how many milliseconds it took. If something is wrong, it shows the reason.

⚠️ If embeddings fail, the service doesn't stop: posts keep going out, but matching falls back to keywords and “meaning-level dedup” stops working — repeats of the same idea become possible. This used to happen silently, so now a notification with the reason arrives in “Notifications” (at most once a day per project). If you see one, press “Check” in the profile — the specifics are there.

💡 The indicator in “Project” → “Feed quality” only shows WHETHER a model is set. The “Check” button answers a different question — whether it works RIGHT NOW.

13YandexART — images 🇷🇺

🔗 Key: the same Yandex Cloud (API key + Folder ID).

🧠 Traits: generation within Russia, tidy illustrations.

✅ Best for: images with data residency in Russia.

14Nano Banana (Google Gemini) — images 🇺🇸

🔗 Key: Google AI Studio (aistudio.google.com/apikey) → «Create API key».

⚙️ Setup: «Google Gemini API key» field.

🧠 Traits: fast, high-quality images, photorealism.

✅ Best for: vivid photo-style posts.

15OpenAI (gpt-image-1) — images 🇺🇸

🔗 Key: platform.openai.com/api-keys.

🧠 Traits: consistent illustrations, good controllability.

✅ Best for: a unified style across a series of images.

16Auto and local mode

How the service picks a model:

  • ✅ «Auto» takes the first provider with an available key.
  • 🧩 With no keys, text is written by a local deterministic rewriter and a gradient placeholder is used instead of an image.

This lets demos and tests run without keys.

Didn't find an answer? Write to us — we'll help.

Contact support