This is the one place to configure your AI model. Changes here apply immediately everywhere — AIK chat, all agents, and all workflows. No restart needed. Keys entered here are personal to your account and always take priority. (Admins also see a separate Platform Keys tab — that one only fills in for other users who haven't set their own key here.)
API Keys — one per provider
Set each provider's key. The right key is picked automatically based on the model you select below — you can add as many as you need.
GPT-4o, o1, o3, o4 models
Claude Sonnet, Opus, Haiku
Gemini 2.0 / 2.5 models
Llama via Groq (ultra-fast)
Off by default. When it is on, the Video Maker asks the service for real footage instead of stock clips, and the run report says what it cost. Veo Lite is the one to start with: it costs about $1.50 for a thirty-second video, needs no key beyond the Gemini one above, and is the only cheap option that makes its own sound. If the service fails or the key is missing, the video still comes out on stock footage and the report says which service refused and why.
Only the one your choice uses. All three Veo tiers need no key of their own — they use the Gemini key above.
Where ComfyUI is listening — on this machine, or another one on your own network.
Build the video workflow once in ComfyUI (Wan 2.2, LTX-2 or CogVideoX), title the positive prompt node AIK_PROMPT, then Export (API) and put the file's path here. AIK writes each scene's line into that node and re-rolls the seed, so the workflow — model, size, length, quality — stays yours. A clip takes minutes rather than seconds, and costs nothing but electricity.
Off by default. Switched on, every run of a Browser agent (Web Surfer) is recorded as a video into My Files → run-videos, named after the step, so you can watch what it did even when it ran on a server with no screen. Long runs make big files — keep only the first minutes if disk matters. The run’s result says where the recording is.
If an agent's model fails because of its provider — the service is down, out of credit, or the model no longer exists — the step tries once more on this model, and the result says which model failed and which answered (the cost is counted on the backup). A task that is simply wrong, or a step that runs out of time, is not handed on. Blank = no backup.
Off: Claude is reached the way every other provider is (the OpenAI-compatible route), which cannot use Anthropic's prompt caching. On: agents on a Claude model use Claude's own API, which caches the instructions and tools an agent resends at every step — in our test the second run of the same agent cost about an eighth of the first. Nothing else changes; untick it to go back. Document extraction and other models are not affected.
On: before a document's text goes to a cloud model, ЕГН, IBANs, e-mail addresses and phone numbers are replaced with placeholders, and the real values are put back into the answer. On photos and scanned PDFs the same values are found on this machine and covered with a grey box before the picture is sent. What it does not cover: a value too blurry to read stays visible on a picture; names are not hidden; e-mails and pages that an agent reads are not masked. A local model is never masked — nothing leaves your machine. For full protection, use a local model.
This does nothing for runs you start here. A run you press Run on in AIK has no Claude session to ask, so it uses the model set above and is billed to that provider — ticking this box does not change that. What it decides is who pays when Claude Desktop or Claude Code starts a run through the connector: ticked, it asks Claude and costs no API credits; unticked, it uses the model above. Script agents run as subprocesses and borrow Claude through a relay, so they follow the same rule.
Status, checked 16 Sep 2026: no Claude app answers this. Claude Code and Claude Desktop both refuse the hand-off (the feature they need was deprecated in the MCP protocol), so a ticked box only makes every run from Claude fail once before it retries with the model above. Leave it unticked; to think with Claude, pick a Claude model above — billed to your Anthropic API key, not your subscription.
Free, instant, no card. With it the Video Maker agent uses real stock clips instead of plain text cards.
With a key, an IF step in a workflow can ask a question in words ("did this step finish the job?") instead of matching text. It is a cloud service — that step's text is sent to typesafe.ai — so leave this empty on an installation where nothing may leave the building.
Large cloud models via Ollama (free plan available — ollama.com)
▸ Custom / other provider (generic key)
Only needed when using a custom base URL not covered by the providers above.
One picker for everything. The model's provider decides cloud vs local automatically — Ollama models run locally (free), the rest use the matching API key.
Asks OpenAI, Anthropic, Google, Groq and Ollama Cloud which models they still run — and calls each one once to be sure. Only models that answer "no longer available" are removed; a model you are using is never touched. Takes a minute or two.
▾ Base URLs (auto-configured — only change if using a proxy)
Use this for OpenRouter, LiteLLM proxy, or any non-standard endpoint.
Smart Auto — which models it tries
Agents set to Smart Auto judge how hard a task is and then take the first model in that row whose API key you have saved. Type model names separated by commas, best first. Leave a row empty to use the built-in order shown under it.
The provider is worked out from the model name, so you only type names. A name with no key saved is skipped; if nothing in the row can be used, the agent falls back to your local model.
Document Extraction (optional)
Your main model is the single AI Model picker above (installed Ollama models appear there after “Fetch local Ollama”). These extra settings only affect document extraction, and only when the model doing it is a local Ollama one — with a cloud model set, extraction uses that model and this card changes nothing.
Document Extraction Settings
Model used specifically for invoice / document extraction. The installed ones are listed after ↻ Fetch installed models (it also runs when the page opens).
Only needed for the llama_index indirect path. Installed models appear after ↻ Fetch installed models.
How to call Ollama during extraction.
Extracting items uses more context and is slower.
Installing or removing local models is done by an admin under Admin → Ollama Models; press ↻ Fetch installed models above to list them here.
Social media tokens (LinkedIn, Slack, etc.) are managed on the Integrations page.
App URL
Your server's web address. Used to create clickable links in notification emails.
Example: http://localhost:8001 or http://192.168.1.5:8001
Outgoing Mail (SMTP)
Used by the Mail Agent to send emails on your behalf.
Incoming Mail (IMAP)
Used by the Mail Agent to read your inbox.
Your own database
The address of a database of yours — your accounting system, your orders, your CRM — so agents can read from it and write into it. For example, an agent can put the invoices it extracted straight into your system instead of into an Excel file. Leave it empty and nothing here is touched: this is not where AIK keeps its own data, and changing it cannot move or break that.
Format: postgresql://username:password@host:port/database — PostgreSQL. It is saved encrypted and belongs to your account alone; nobody else on this installation can see or use it.
Which set of tables to use inside that database. Leave blank unless someone told you otherwise.
⚙️ These are the site's shared defaults, not your personal keys. Every user has their own keys on the AI / LLM tab — those always come first. The keys you set here are only used as a fallback, for any user who hasn't entered their own. Same storage as the Admin page.
Platform API Keys & Environment
Saved straight to the database and used immediately — no restart needed. To remove a saved key, click its Clear button, then Save.

Loading…

🎙️ Voice Agent — pick the language, and choose the engine for the ears (listening) and mouth (speaking) separately. Mix them freely. Language & engine changes take effect on the next call. Changing the free STT model needs a server restart. The Cloud options need their API keys set by an admin in Platform Keys (Deepgram / ElevenLabs); without a key they safely fall back. Note: every call still uses your Twilio phone minutes, and the brain (LLM) is billed per your AI/LLM model choice.
🌍 Language
The language the agent greets in, listens for, and replies in.
👂 Ears — understanding (speech-to-text)
How the agent understands the caller. Twilio understands Bulgarian best (small per-call fee); Free is fully self-hosted (only phone minutes).
Free STT model — only used when ears = Free:
🗣️ Mouth — voice (text-to-speech)
How the agent speaks. Free = self-hosted Piper (nicer Bulgarian voice); Twilio = built-in voices (small fee).
🧩 Voice Agent Template — what it does
Pick a ready-made role to start from, then edit the instructions to fit your business. The agent still speaks in the language above.
Leave empty to use the default friendly assistant.
⚠️ Advanced Developer Features — CLI Tool
Allows admins to create CLI Tool agents (Python scripts that run on this server). On-premises installations only.
🔒 Not available — the platform owner has not granted this permission yet.
🗑️ Data Retention
Automatically delete old chat history after a set time. Saved facts & preferences and the tamper-proof audit log are never auto-deleted. Default: keep forever.
The step-by-step details of old runs are the biggest thing the database holds. The runs themselves — status, model, cost, output — are kept for ever, so history, ROI and billing stay whole.
📈 ROI Indicator
Powers the dashboard "Time saved counter". Set how long each task would take a human by hand; add your labour cost per hour to also show money saved. (Per-workflow overrides are set on each workflow.)
📜 Action Certificates
Certificates are digitally signed by this install. Publish the public key below so anyone can confirm a certificate genuinely came from you. Point the "Verify" link on certificates at a public URL (e.g. your website) so clients can click it — verification is self-contained, so any AIK verifier checks any certificate.
🔄 Keep your platform up to date. On the cloud, updates apply automatically. On your own (on-premise) install, click below to pull the latest features — no terminal needed.
🔑 License
Your product license. Updates require a valid license. Paste the signed key you were given.
Checking license…
Update source
Where updates come from. Leave blank to update from GitHub (needs repo access). Set your signed update server (e.g. https://aikautomation.com/updates) to update from sealed packages — no repo access.
Platform Version & Updates
Check for newer versions and apply the update automatically — no terminal needed.
Reading the installed version…
↩️ Roll back
Undo the last update if something went wrong — restores the app and its database to exactly how they were just before the update, then restarts. (Anything created after the update is rolled back with it.)
Checking…
🧩 MCP servers
The tool servers the AI uses. Test all actually starts each one and lists its tools — a quick way to confirm nothing broke after an update.
Loading…
🔐 Two-Factor Authentication (2FA)
Add an extra layer of security to your account. After enabling, you will need a 6-digit code from your Authenticator app every time you log in.
Checking…
📦 Your data
Download everything this platform holds about your account — your agents, workflows, settings, run history, saved facts and every file in your sandbox — as one zip. Passwords and stored API keys are left out on purpose: they stay readable in Settings, and a secret inside a downloaded file is a secret that can be lost. (GDPR, right of access.)
⚠️ Danger Zone
Permanently delete your account and all your data. This cannot be undone.
📁 Folders outside AIK
Your agents work only with your own files (My Files). Switch this on to let them also use folders outside AIK — for example a shared invoices folder on this computer or the network.
You switch this on yourself and you choose the folders. Your agents will be able to read, change and delete files in them, so you are responsible for what you open. AIK’s own files, other users’ folders and a whole drive (like C:\) are always refused.
🗄️ Database structure changes
The SQL agent can always read your databases, add rows, add a column and create a table. Switch this on to let it also rename or delete columns and tables. Each change still waits for your OK, and your own database.db is copied to My Files/.backup first, so a delete can be undone. AIK's own database is never touched.
🖥️ Self-hosted installations only. These instructions are for running AIK on your own machine (Windows, Mac, or Linux). If you're using the cloud version at aikautomation.com, skip this tab — everything is already set up for you.
🚀 Install AIK on Windows
One-click installer — sets up everything automatically: Python environment, database, Windows service, and a Desktop shortcut.
Step 1 — Get the AIK files
Clone the repository from GitHub, or unzip the installer package you received.
Step 2 — Double-click setup.bat
Windows will ask "Allow this app to make changes?" — click Yes. The installer runs automatically.
✓ That's it!
The installer creates everything automatically:
• Python environment + all dependencies
• Database (SQLite, ready to use)
• Windows service (auto-starts on boot)
• Desktop shortcut (click to open AIK)

First login: admin@local / the password printed once at the end of the install — write it down
Change the password after first login in Settings → Security.

After install: open Settings → 🔄 Updates and paste your license key. Updates are already pointed at the official source — turn on "Install updates automatically" if you want them hands-free.
No prerequisites. The installer sets up Python 3.12 automatically if it's missing — the client installs nothing beforehand, just double-clicks setup.bat. If the provider baked a license into the package, it's applied automatically too.