How to get an AI API key and connect it in Forgebot

We don't sell tokens or hand out our own keys: you pay the provider directly at its prices. Customer messages in this block go to the provider you choose.

To try it for free, use OpenRouter (a model with ":free" in its id) or Gemini. For steady use with real customers, use any paid key with a spending limit.

• Sign up at platform.claude.com (Claude Console).
• Top up: Settings → Billing → Buy credits. The API does not work without credits — Anthropic's own docs say to "buy credits before you use the API". Credits expire one year after purchase.
• Open Settings → API keys (platform.claude.com/settings/keys) → Create key.
• Name the key, set an expiration, and make sure to pick a SINGLE workspace in the Workspace field. A key that works across all workspaces needs an anthropic-workspace-id header on every request — Forgebot does not send it, so such a key will be rejected.
• Copy the key — shown once, starts with sk-ant-.

Spending limit: Settings → Billing → Spend limits. Each account tier has its own monthly cap; hitting it means waiting for the next month or asking for a raise.

• Sign up at platform.openai.com.
• Create a key: platform.openai.com/api-keys.
• Add credits: Settings → Organization → Billing.
• Copy the key in full.

Tiers: Free (low limits, not available in every country), Build after $5 in total purchases, Launch from $100, Grow from $500 (checked 2026-10-11). Hard spending cap: platform.openai.com/settings/organization/limits.

• Open Google AI Studio and sign in with a Google account (18+).
• New users get a project and key created automatically after accepting the terms (confirmed 2026-10-11). Keys live on the API Keys page; need another one — Create API key.
• For higher limits, press Set up billing and prepay at least $5.

There is a free tier, but with low limits. Google warns that on the free tier your prompts and responses may be used to improve Google products — use the paid tier for customer conversations.

Old unrestricted keys are rejected by the provider: keys created from 2026-05-28 are already the new "auth key" type — if an old key stops working, create a new one.

• Sign up at platform.deepseek.com.
• Top up: platform.deepseek.com/top_up (PayPal, card, Alipay, WeChat Pay).
• Create a key: platform.deepseek.com/api_keys.
• Copy the key.

Your balance doesn't expire, and an unused balance can be refunded (Billing → Refunds). Out of money — the provider returns an insufficient-balance error. Servers are in China — Forgebot shows a warning under the model picker.

• Create an Alibaba Cloud international account: email, password, country (cannot be changed later), payment details.
• Activate Model Studio and pick the Singapore region — the one we recommend.
• In the Model Studio console: API Key → Create API Key — default workspace, All permissions.
• Copy the key (starts with sk-) — it is shown only once.
• Find your Workspace ID on the Workspace Management page: for the Singapore region Forgebot asks for it when you connect the key.

A key is locked to its region: a key from one region does not work in another. Forgebot accepts keys from two regions only — Singapore (recommended) and US (Virginia; not confirmed to work with Forgebot). Keys from China (Beijing), Germany (Frankfurt), Japan (Tokyo) and Hong Kong are not accepted by Forgebot — you will not be able to connect them.

Why the US is "not confirmed": as of 2026-10-11 Alibaba's own documentation contradicts itself — the regions table marks the address Forgebot uses for the US as not supported, while the endpoint reference lists it as still available. We have not tested this region with a live key. The 90-day free quota for new users is also Singapore-only — pick Singapore.

• Sign up at console.x.ai.
• Add credits: the API does not work without them (confirmed 2026-10-11).
• Create a key on the API Keys page in the console (console.x.ai → API Keys).
• Copy the key.

xAI only spends prepaid credits: once they run out, requests are rejected. Auto top-up can be enabled from $5 per charge, with a monthly cap.

• Open console.mistral.ai and sign up.
• On the left — API Keys → Create new key → a name and an expiration → Create new key.
• Copy the key — shown once.

Free mode gives API access with no card, with a monthly allowance (confirmed 2026-10-11: "API access is enabled by default with no credit card required"). For more, turn on pay-as-you-go in Admin Panel → Subscription. Servers are in the EU.

• Sign up at platform.kimi.ai.
• Top up: at least $1 to start. Individuals can only pay via WeChat Pay and Alipay (QR code, confirmed 2026-10-11); businesses go through sales.
• Create a key: platform.kimi.ai/console/api-keys.
• Copy the key.

For Ukraine, Kazakhstan and most CIS countries, paying via WeChat Pay/Alipay is almost certainly a dead end — weigh that before picking Kimi. A key from another regional Kimi platform is rejected. Servers are in China.

• Sign up at openrouter.ai.
• Create a key: openrouter.ai/settings/keys — give it a name and a credit limit (we recommend one).
• Copy the key — starts with sk-or-v1-.
• For paid models, add credits: openrouter.ai/settings/credits (cards, AliPay, crypto).

Free (":free") models: 20 requests per minute for everyone; 50 per day with under $10 in lifetime purchases, and 1,000 per day after $10+ in purchases (confirmed 2026-10-11).

• Open Settings → the "AI Keys" tab.
• Press Connect next to the provider you need.
• Paste the key into the API Key field (for Qwen, also pick the Key Region — Singapore or US; for Singapore, enter the Workspace ID).
• Press Save Key — it is stored encrypted, only the last 4 characters stay visible.
• Press Verify.

For Anthropic, OpenAI, Gemini, DeepSeek and Mistral, Verify reads the free model list — "Working" means "key accepted", not "there is money on the account". For xAI, Qwen and Kimi it sends a tiny model request. For OpenRouter it reads the key's remaining limit.

The Connect / Replace / Delete buttons are visible only to someone with permission to manage AI keys.

Add the AI-assistant block to the flow and open its panel. The fields below follow the panel order and use the same names.

• What the Assistant Does — "Replies to One Message" (one reply, then the flow moves on; the block is marked "one reply" on the canvas) or "Holds a Conversation Until the Goal Is Met" (the assistant answers every customer message until it reaches the goal you set; "conversation" on the canvas).
• When the Goal Is Met — "conversation" mode only: the goal, up to 1,000 characters. Without it, such a block cannot be published.
• Role & instructions — up to 4,000 characters: who the assistant is and how it replies. The customer never sees it.
• Knowledge base — up to 20,000 characters: facts (menu, prices, hours, terms). Keep customer personal data out of it.
• Provider — only providers with a connected key are listed.
• Model — "Curated Models" (ones we tested) or "From Your Provider Account" (a live list from your own key).
• AI Reply Limit per Person — optional: a number and a period (per minute, per hour or per day). On the free plan Forgebot already caps it at 20 replies a day per person (the field can only lower it); on a paid plan there is no cap by default, but we recommend setting one — otherwise a single person could spend as much as they want on your key.
• New Conversation After Silence (min) — shown in "conversation" mode and when memory is on: 1 to 1,440, 30 by default. If the person stays silent that long, the assistant forgets the previous conversation, and in "conversation" mode the conversation ends via "silent/limit". With memory on, there is also an "Also on /start" switch (on by default). The reply limit is not reset by this.
• Advanced — a collapsed section: "Remember the Chat Between Flow Runs" — the assistant remembers the conversation with this person even if the flow restarts (2 to 50 recent messages); "Max AI Replies per Conversation" — "conversation" mode only, up to 25 replies, 20 by default; "AI Reply" — "Send to the Customer" or "Save to a Variable Only" (the customer won't see the reply, it goes to {{ai_reply}}; not available in "conversation" mode); "If the AI Can't Reply, Message the Customer" — up to 1,000 characters; "Max reply length (characters)" — 50 to 3,500, 800 by default.

In "one reply" mode the block has three outputs:

In "conversation" mode there are four outputs:

Connect "error" to a "We'll get back to you shortly" message and/or an operator. If it is left unconnected, the bot continues on its own as if the AI stayed silent: in "one reply" mode via "replied", in "conversation" mode via "silent/limit". The customer then gets only the text from "If the AI Can't Reply, Message the Customer" (in Advanced), or nothing if that field is empty — so connecting "error" is worth doing.

The reason code is in {{last_ai_error.reason}} — you can branch the flow on it (for example, "no credit" → notify yourself, everything else → "we'll get back to you").

/start and a new conversation after silence do NOT reset the reply limit — only the assistant's memory resets. The provider's own limits (per minute/day/month) apply on top, visible in its own dashboard.

• The provider refuses → the block takes the "error" output, reason no_credit.
• AI Keys shows "No Credit", and the block panel shows a warning.
• To avoid hammering the provider, the block skips it for 5 minutes and goes straight to "error", then tries again.
• Top up → press Verify → it recovers on its own.

A rejected key (deleted, expired, wrong) isn't used again until you replace it or press Verify.

The code lives in {{last_ai_error.reason}} — use it in a Filter block condition to handle different reasons differently.

• Set a spending limit at the provider — it's your fuse against a surprise bill.
• Connect the "error" output to a "We'll get back to you shortly" message and to an operator.
• Start with a cheap, fast model from "Curated Models"; switch to an expensive one only if replies disappoint.
• Put facts in the knowledge base, not in the role: the role is "how to talk", the base is "what to say".
• Don't store customer personal data in the role or the knowledge base.
• For stricter privacy, pick providers with servers in the US or EU: DeepSeek and Kimi run in China.