How do they work together?

A simple real-world example: you want to automatically process incoming invoices every morning.

Step by step

Choosing a backend type

When creating an agent from scratch (+ New Agent), you pick how it connects to its tools. There are three options:

Using the template library

Don't want to start from scratch? The Agents page has a built-in template library with ready-made agents for the most common tasks. Just click a template, give it a name, and deploy it in one click.

Advanced agent types

Beyond the basic workbenches, the platform includes specialist agents for powerful automation scenarios.

Agent Knowledge Base (Memory)

Every agent has a personal knowledge base — a set of key/value facts it can remember between runs. Open an agent and click the Knowledge Base tab. Click + Add Entry, type a key (like "default_recipient") and a value (like "boss@company.com"), then save. The agent can read and update these entries during its runs — useful for storing settings, preferences, or accumulated knowledge.

Agent Self-Improvement (opt-in)

Each agent can review its own past run logs and automatically suggest improvements to its instructions. Open an agent, go to the Self-Improvement tab, enable it, and pick how often to run analysis. The agent will periodically read its recent results and rewrite its own system message to perform better on future tasks — no manual tuning needed.

Your board

When no agent is open, the middle of the Agents page is your board. Drag the agents you use most onto it from the list on the left, then drag the cards into any order; ✕ takes one off, and a click opens it. The Workflows page has its own board. Both are saved to your account, so they look the same on every computer and phone.

Keyboard shortcut & output saving

Step by step

IF Conditions — branching your workflow

Sometimes you want the workflow to take a different path depending on what happened in a previous step. For example: "If the email contains the word 'urgent', notify me immediately. Otherwise, just log it."

Asking a question instead of matching words — the decide option

The four rules above look for text. Sometimes the question is not a word but a judgement: “did this step actually finish the job?”, “does this reply sound like a complaint?”. With a decision-model key set, the IF step offers a fifth option, decide (AI, yes/no): instead of a word you type the question, and the model answers yes or no with a confidence, which the run log records next to the branch it took (“TRUE ✓ (confidence 87%)”).

Cost: one condition is a few hundred tokens — a fraction of a cent — billed to your own account; AIK adds nothing. The step records its tokens and its cost in the run history like every other step — the Runs page's Cost (your key) column, which counts only what a paid model charged on your own key (steps on a local model add nothing); if you go through a gateway, that gateway's own activity page is the exact bill.

The Canvas view

Click the ⬜ Canvas button in the workflow header to see a visual map of your workflow. Each step is a card, connected by arrows. IF steps show two output arrows — one green (TRUE) and one red (FALSE). Human Approval steps appear as a purple 👤 node. You can double-click any card to edit that step.

Knowledge for a single step

An agent's card has a Knowledge Base — facts it is given on every run. A workflow step gets those whenever it is linked to a saved agent (the Use a saved agent field in the step dialog). A step that is not linked to one has its own box, Knowledge for this step: type plain lines — “Our VAT number is BG123456789”, “Only invoices over 2000 need approval” — and they are handed to the step every time it runs. A linked step gets both, its agent's facts and the step's own.

Under the box there is Learn from runs. Tick it and the step fills the box itself: after each successful run it writes down what it found — an address it discovered, a file name, a pattern — and replaces a fact it already had rather than repeating it. The box keeps the last 40 lines, and you can edit or delete anything in it. A step linked to a saved agent learns into that agent's Knowledge Base instead, exactly as before.

👤 Human Approval — pause and wait for a person

Add a Human Approval step anywhere in your workflow to pause execution and wait for a real person to review and decide before continuing. This is useful for high-stakes actions — like approving a payment, publishing content, or sending a bulk email — where you don't want full automation.

Trigger Chains — one workflow starts another

Use a Sub-Workflow step to run a second workflow as part of the current one. This lets you build modular automation — for example, a "daily report" workflow that calls a "send email" workflow as its last step, keeping each piece reusable on its own.

Or set When this succeeds, run in the workflow's settings: after every successful run — by hand, on schedule, from an e-mail, a webhook or after your approval — AIK starts the next workflow itself and hands it the result. It starts only when the run really succeeded (not failed, not waiting for your OK), only if the next workflow is switched on, and at most five in a row, so two workflows can never keep starting each other.

Step shortcuts — duplicate & single-step run

Each step in a workflow has two extra action buttons next to the usual edit/delete:

Failure notifications

The Failure Notifications toggle in the Schedule section of any workflow controls whether you receive an email when a scheduled run fails. When On, the platform sends a failure email to the address configured in your mail settings. This is a global setting — it applies to all your scheduled workflows, not just the one you're looking at.

Below the toggle there is a Send this workflow's failures to box. Put an address there and that workflow emails it whenever a run fails — manual or scheduled — even if the toggle above is off. Useful when one workflow matters to a colleague who does not have an account. Click Save settings to store it.

Dry Run — test without side effects

Enable Dry Run mode in the Run panel before clicking ▶ Run. The workflow executes all its logic but skips any actions that would have real-world effects (sending emails, posting to social media, writing files). Use it to verify the flow and see what the agents would do before committing.

Smart Recovery — automatic retry on failure

Set Max Retries on any step (in the step settings). If a step fails, the runner automatically tries again — up to the limit you set — before marking it as failed. Combine with a Timeout (seconds) to prevent a hung step from blocking the whole workflow.

Time-Travel Replay — restart from any step

If a workflow fails halfway through, you don't have to start all over. Open the Analytics page, click Debug on any run, and in the step timeline click Replay from Step N next to the step where things went wrong. The workflow restarts from that step only — all steps before it are skipped, saving time and cost.

Import & Export

You can save any workflow as a file and reload it later — useful for backups, sharing with a colleague, or moving workflows between environments (for example, from your local machine to the cloud server).

How to use it

The two money figures

They are deliberately separate, because they come from different places.

❓ "What I couldn't work out"

Every report ends with the questions AIK couldn't answer from your description — which mailbox folder, what happens when a total doesn't match. This is the most valuable section. It is exactly where automation goes wrong, and it is your list of questions for the next conversation with the client.

Sharing the report

Letting AIK build it

Tick "Also build this workflow" and AIK creates the workflow it described, with the steps it identified.

How it works

Three ways to run a pipeline

🧪 Trial — let it prove itself first

Before a new document type goes to work for real, it can work beside the person who does the job today. Tick Trial at the bottom of the type's Edit form, name the spreadsheet that person already keeps (upload it to My Files), choose the field both sides share — usually the invoice number — and how many correct answers in a row it must reach.

🎯 Test models — which model reads these best?

Before you switch to another model, measure it in minutes. Put a few sample documents in a folder, and a spreadsheet with their correct values in My Files (the same kind of sheet the trial uses). Press 🎯 Test models on the type, pick the folder, the spreadsheet, the field both share (usually the invoice number) and up to three models. Each model reads every document — nothing is saved, sent or moved — and you see per model how many fields it got right, every difference, how long it took and what it cost on your key. Up to 20 documents per test; formats do not count as mistakes, exactly as in the trial.

The built-in Invoice type

"Invoice" comes ready-made: vendor, invoice number, date, total amount, line items — and, optional, net amount, VAT amount and the supplier's EIK for the VAT and EIK checks — with results going to the database and Invoices.xlsx, exactly like the classic invoice pipeline. Its math, VAT, EIK, IBAN-change and “unsure” checks are on by default: an invoice that fails one waits for your OK instead of being saved. On the Dashboard (and in the AIK chat, when AIK ran it) each held invoice has Approve, Refuse and 📄 Open file to see the document itself. A refused invoice is moved from Processed to Failed and nothing is written. A check's message shows its numbers — e.g. what 20 % VAT on the net would be and the total that gives, next to the invoice's total. Where the documents come from: My Files, or a folder outside AIK only after you open it under Settings → Security → Folders outside AIK — then it appears in the list. Press Edit on a built-in type to see and change all of it: your first save quietly makes your own copy with the same name, which replaces the built-in on your list and for AIK — other accounts are not affected (records from before stay with the built-in). Customize makes an extra copy — every press makes a new one (“Invoice copy”, “Invoice copy 2”…), which opens for editing so you can name it straight away.

More options on each type

When something goes wrong

One bad file never stops the batch. Files that can't be read are listed as failed with the reason; files with wrong values are invalid with the exact check that failed; records caught by the risk, math, VAT, EIK, new-vendor or strict-signature checks are flagged and wait for your approval — each with a plain-language reason. A wrongly extracted record can be deleted with the 🗑 button under Records, so the document can be processed again. Every run leaves a "Last run" status line on the type's card (green = OK, red = the error text), and the full history is on the Runs page. Need to abort a long run? Press ⏹ Stop on the type's row — the pipeline halts after the current document and reports how far it got.

AIK on your phone: on Android open aikautomation.com in Chrome → ⋮ → Install app. AIK gets its own icon, opens full-screen, and appears in the phone's Share menu — share a photo or a PDF to AIK and it is saved in My Files → from-phone. On an iPhone: Safari → Share → Add to Home Screen (iPhones do not offer the Share menu to web apps).

The Orb and the AIK Chat page are the same AIK — same answers, same language, same tools. AIK answers in the language you write in — Bulgarian or English —; when something it started needs your OK, the answer ends with an Approve / Refuse card — the same decision as on the Dashboard, and the card shows the result (or the decision you took on the Dashboard) — and the Orb shows the provider and model you chose in Settings (change them there). On a phone, the 🕘 button next to the typing box opens your past chats. AIK keeps its answers short. It can see which services you connected on Integrations, create the right agent for each (a LinkedIn agent uses your LinkedIn connection — it never asks for your password), make its built-in tools (Video Maker, Video Editor, Website Maker, CV Maker, Invoice Maker, Meeting Notes, Document Converter), and run your agents and workflows — each run appears on the Runs page, counts toward your budget and waits for your OK where you asked it to. It can also look at your tickets (admins), analytics, notifications, assistants, the Marketplace, schedules and the agent fence; create and change Assistants, schedules and your alerts; install a Marketplace template; and open a ticket. A ticket reply from AIK is only a draft — it goes out when a person presses Send in Support.

Support page: how-to questions go to AIK chat (it reads this guide); something broken or a request for a person goes in a ticket. Admins see the Ticket inbox there: reply, close, delete a ticket, tick several and press Delete selected, or Delete old closed (closed tickets older than 90 days — open ones are never deleted). Refresh says how many tickets there are; the Super Admin's AI drafts: on/off label leads to the switch on Super Admin.

What can you ask AIK?

Tips for better results

How it's built — ears, brain, mouth

Pick your combination (Settings → Voice)

Ears and mouth are chosen separately, so you can mix them — e.g. Twilio ears + free Piper mouth is accurate and cheap. Pick the language (Bulgarian or English), and choose a Voice Agent Template (Receptionist, Customer Support, Appointment Booking, Sales…) then edit the instructions to fit your business.

Setting up Twilio (the phone line)

Twilio connects your agent to the real phone network. You need a free account, a phone number, and a few keys — all pasted into Settings → Platform Keys. Saved values there show blurred — click Show to read one; passwords and API keys only say they are saved.

Optional: Cloud voice keys (most accurate Bulgarian)

Only needed if you pick the Cloud ears/mouth. Both have free trials, and if a key is missing the agent safely falls back. Paste keys into Settings → Platform Keys.

On-premise / self-hosted

To use the Free engine on your own machine you need the voice engines — the installer fetches them for you (Windows, Linux and macOS alike). Only if the install summary listed them as missing, fetch them again with python scripts/install_voice_engines.py. This downloads the Piper voice binary and the Bulgarian + English voices for your OS. Whisper installs automatically with the app. Then the free ears+mouth work exactly like on the cloud.

The kinds of memory

You're always in control

Make one in a minute

The channels

Appointment Booker — the ready-made pipeline

The built-in Appointment Booker runs the full flow: it checks your Google Calendar or a plain Excel schedule for free times, offers the caller slots, confirms, books it — and can call the customer later to remind them. Point it at your calendar or an .xlsx in your files, and it works on chat and phone alike.

Invoicing in the menu has three tabs. Your company: name, ЕИК, VAT number, address, МОЛ, IBAN — printed on every invoice; set the first number once if you already issue invoices elsewhere; untick VAT registration and invoices go out without VAT and say so. Customers: type them in, import an Excel or CSV with columns like Фирма, ЕИК, ДДС №, Адрес, МОЛ, or type just an ЕИК and press Look up to fill the name and VAT number from the Trade Register and VIES. Invoices: pick a customer, add lines, choose 20, 9 or 0 % (0 % needs the legal ground as a note) and issue. Numbers are 10 digits and follow each other with no gaps or repeats; VAT and totals are calculated by the program, not by AI; the PDF is saved in My Files → invoices. It carries no visible AI mark — it is made by code — but a hidden signature that locks the IBAN and the total, so a changed copy no longer verifies. The Invoice Maker agent does the same from plain words (“Invoice Клиент ООД for 10 hours of consulting at 50 € each”); the customer must already be on the Customers tab. Many at once: on the Invoices tab upload an Excel or CSV with one row per line — Клиент (name or ЕИК), Описание, Количество, Цена, and optionally ДДС %, Мярка, Бележка, Фактура. Rows for the same customer and VAT rate become one invoice. You see a preview first; if a row names an unknown customer or has a bad number, nothing is issued until it is fixed; Issue all then numbers them one after another. The agent does the same with “issue the invoices from billing.xlsx” (a sheet in My Files) or with several customers in one request. Nothing is sent by itself. Report: Analytics → Invoices — issued and read shows, for any period, the invoices you issued (sales) and the ones AIK read for you (purchases): count, net, VAT, total, top customers and suppliers, month by month, and Excel for my accountant downloads both as two sheets laid out like the VAT sales and purchase ledgers. Keeping them: Invoicing → Your company → Keep invoices for (5 years unless you change it — check the period with your accountant). AIK never deletes invoices by itself; deleting your account while invoices are inside that period asks you first to download them. Unpaid invoices: every invoice gets a due date (the invoice date plus the Payment term on the Company tab, 14 days unless you change it). Mark paid on the Invoices tab records the payment. Switch on E-mail me when an invoice is past due and every morning AIK checks: when an invoice has become late, you get one e-mail listing the late ones — never the same list twice.

Your company's own database (e.g. PostgreSQL on an on-premise install): give the SQL agent its address — agent → Environments, or Integrations → Other services, name SQL_CONNECTION_STRING, e.g. postgresql://user:password@host:5432/erp. You do not need to know the columns: the SQL agent first reads the database's own description (inspect_schema, get_table_info) and then writes its query or new row. It can also add a column or create a new table; renaming or deleting a column or a table is off until you switch on Settings → Security → Database structure changes — then each change still waits for your OK, and your database.db is copied to My Files/.backup first. Every insert, change, new column or new table waits for your OK when Ask before acting is on. AIK's own database is always refused.

Super Admin — AIK's own database: ask AIK Chat, e.g. "how many runs failed this week, per account?". It can only read (one SELECT at a time, at most 200 rows), never sees passwords, keys or tokens, and every question is written to the activity log.

Your own database: every account has one — database.db in My Files (a small SQLite file). The SQL agent and the analysis agent read and write it, AIK Chat answers questions about it ("show my database", "how many invoices this month?"), and a document type saves its records there when you write my database under Also insert into another database. A database file elsewhere on the server, or AIK's own database, is refused.

Three ways to work with files

pip install websockets requests pystray pillow
python aik_worker.py --tray --folders "C:\Invoices,D:\Reports"

No empty folders left behind: a folder an agent or AIK made while working and left empty is removed when the work ends. Your files, folders with something in them, and a folder you asked for by name are never touched.

A file on your PC, used by any agent: name the file in the task (“upload demo.mp4” — a name with spaces works too). If it is not in My Files but is in a folder you shared with the Worker, AIK copies it into My Files → from-pc before the agent starts, and the run report says so; if it is in neither place, the run report says that instead. Up to 10 MB per file; nothing is written back to your PC. The ⬆ Sync button on a Worker file does the same by hand. Videos, PDFs and pictures need Worker 2.2 or newer — older Workers sent every file as text, which broke them; AIK now refuses rather than save a broken copy, so download the new Worker once.

Local Worker capabilities

When the worker is running, AI agents get these extra abilities on your machine:

Quick start (cloud users)

If you're using AIK Platform in the cloud, here's all you need to do:

The worker starts automatically with Windows and runs in the background. Already installed? Just start AIK-Worker.exe — it reconnects instantly. The ↺ Refresh button on My Files checks the connection immediately.

Advanced usage

# With a local SQLite database
python aik_worker.py --tray --db sales:sqlite:C:\data\sales.db

# With a custom MCP tool server
python aik_worker.py --tray --mcp mytools:python:C:\tools\my_server.py

# Re-run setup wizard
python aik_worker.py --setup

Simple schedule (time + days)

The easiest way. Open any agent or workflow, scroll to the Schedule section, enable it, pick a time and tick the days you want. Done.

Cron expression (advanced)

For more precise control — like "every 30 minutes" or "on the first day of every month" — switch to ⚙ Cron mode and type a cron expression.

When an e-mail arrives

An agent, a workflow or a document type can start by itself when a new e-mail lands — an invoice from a supplier, an order from a client. Tick Run when an e-mail arrives on the agent's Run tab (then Save Changes) or in the document type editor, and optionally say who it must come from and which words the subject must have (any of them, separated by commas).

• It uses the mailbox you set in Settings → Mail and checks it every minute.
• Only e-mails that arrive after you switch it on count — your old inbox is never processed.
• Each e-mail starts one run, once. Nothing is marked as read, so your mailbox looks the same as before.
• Attachments are saved in My Files → inbox/<date>. A document type reads them; an agent gets who sent the e-mail, the subject, its text and the attachments' names together with its task.
• If the mailbox cannot be opened (a changed password, for example), one failed run says why in the run history.
• Mailbox: Settings → Mail (address + app password), or Microsoft 365 — the account connected with 🔗 Connect on Integrations, for company mailboxes where the address-and-password way is switched off.

Watchdog — only tell me when something matters

Tick Watchdog on any agent's Run tab and choose how to be told: E-mail (to your account address) or Telegram (your bot, with a default chat set on the Integrations page). On a scheduled run, or one started by an e-mail, the agent answers "NOTHING TO REPORT" when nothing needs you — then nothing is sent and the run history just says "Nothing to report." Anything else is sent to you, marked as AI-written. If the message cannot be sent, the run says so instead of staying quiet. Too chatty? Write in Only tell me when… what really matters to you (“a client e-mail has waited more than 2 days”) — it is kept on the agent and used on every run.

Ready-made: the Mail Watchdog template checks your inbox every weekday at 09:00 and tells you about e-mails from people that have waited more than 2 days for a reply.

How it works

API keys let external apps talk directly to your AIK platform. Add the key as a header on every request — that's it.

Available endpoints

POST /v1/chat
{"message": "Summarise the latest invoice"}

GET /v1/workflows

POST /v1/workflows/{id}/run
{"variables": {"recipient_email": "ceo@company.com"}}

Security tips

Setting it up

"aik": {
  "command": "cmd",
  "args": ["/c", "npx", "-y", "mcp-remote",
           "https://your-aik-address/mcp",
           "--header", "Authorization:${AUTH_HEADER}"],
  "env": { "AUTH_HEADER": "Bearer YOUR_KEY" }
}

claude mcp add --scope user --transport http aik https://your-aik-address/mcp \
  --header "Authorization: Bearer YOUR_KEY"

Two AIKs at once

If you use both a hosted AIK and one on your own machine, add them under different names — Claude keeps one entry per name, and a second aik is quietly ignored while the first one keeps answering. The Connect Claude box does this for you: a local install writes aiklocal, a real address writes aik. If Claude lists agents you do not recognise, it is talking to the other one — claude mcp list shows which. And the key is fixed once added, so changing it means claude mcp remove aiklocal --scope user and adding it again.

The Claude app’s own Connectors list

Claude (and Cowork) can also take AIK under Customize → Connectors → Add custom connector: the address, authentication “No sign-in”, and the key under Request headers as x-api-key. This route only works for an AIK with a public web address — those calls come from Anthropic’s servers, not from your computer, so localhost can never be reached that way. For an AIK on your own machine use Claude Code, as above. One added there also shows up in Claude Code, which is the usual reason two installs get confused.

When it says the key is missing

Two different messages, two different problems. “X-API-Key header required” means no key arrived at all — the key was never saved with the connection, or Claude is calling a different AIK. “Invalid or revoked API key” means one arrived and AIK refused it. The Last used column on the API Keys page settles it: still Never after Claude has tried, and that key is not the one being sent.

New AI models

Settings → AI/LLM → Check providers for new models asks every provider what it offers now and tests each one with a tiny call. It shows how far it has got — “42 of 130 models checked” — and takes a few minutes. Press Apply and the new models appear in the picker straight away; no restart. In the report, BROKEN means the model was kept: the provider still lists it, our test call failed, and the error was not proof it is gone.

Language

The picker at the top of the app switches between English and Bulgarian. It changes the navigation and the translated parts of the app, and it also sets the language AIK replies in. Pages that have not been translated yet stay in English — more of them are being marked up over time.

Who pays for the thinking

Talking to Claude costs you nothing extra — it is your Claude plan. When you ask it to run something, you choose whose model does the work.

Status, checked 16 September 2026: neither Claude Code nor Claude Desktop currently agrees to do the thinking for a run — the hand-off they would need was deprecated in the MCP protocol in July 2026. A run that asks fails at once with a clear message and Claude re-runs it with AIK’s model. Press Check status beside the box to see what the last Claude app that connected declared — it says whether that app offers to do the thinking, without spending a run. Until the apps support it again, leave the box unticked; to have Claude think, set a Claude model in Settings → AI/LLM, which is billed per token to your Anthropic API key rather than to your subscription.

What Claude can and cannot do

What you can see

Stop a running job

If a workflow is stuck or running too long, click the red ⏹ Stop button in the Live section. The current step is interrupted and the run is marked as stopped. Stop cannot take back what was already done — the run then says exactly what had gone out before you pressed it (an e-mail, a LinkedIn post…), so you do not run it again and send it twice.

When an agent fails, it says why: if the AI model ends without an answer, the run lists the steps it took and what the last one showed, instead of a bare error. A Web Surfer that fails on a mailbox website (abv.bg, Gmail…) also tells you the reliable way: connect the mailbox to the mail tool by IMAP (an Email agent → Environment Variables → Email preset), which reads it directly and saves attachments into My Files. And if AIK ever answers "I'm starting now" without doing anything, it is made to act at once or say what stops it.

📈 ROI Indicator — "Time saved counter"

At the top of Analytics, a green card estimates the human time your automations saved — this month and all-time. Each successful run counts for a "manual minutes" figure (how long that task would take a person by hand), so you see real ROI instead of just token counts. Set it up in Settings → ⚡ Advanced → ROI Indicator: a default minutes-per-run and, optionally, your hourly labour rate to also show money saved (e.g. "142 hrs saved this month · ~€4,200"). Each workflow can override the default in its own settings (⏱ Manual minutes / run), so different jobs count for their real effort. The card shows minutes when it's under an hour, hours above that, and the figure is also included in the weekly report email.

📋 Accountability Report

One click turns your tamper-proof activity log into a plain-English report a manager or auditor can read — grouped by day, with the live integrity check and any flagged items. Open it as a printable page or download a PDF from the buttons on the Analytics page. Because the log is SHA-256 hash-chained, the report can state whether anything was altered (EU AI Act Art. 12 record-keeping). Admin only.

📜 Action Certificate

On any run in the history, click the 📜 button to get a tamper-proof receipt for that action — what the agent did, when, on whose behalf, within what limits, who approved it, and a verifiable integrity seal (the run is recorded in AIK's SHA-256 hash-chained log, EU AI Act Art. 12). Open it as a printable page or save a PDF. Every certificate is digitally signed — anyone can confirm it's genuine at /certificate/verify, offline, with the public key printed on it, without logging in or touching your database. It turns an invisible AI action into something you can show a manager, client, or auditor — provable, not "trust me".

🪪 Agent Passport

A certificate proves one action; a passport is the agent's whole ID card + track record in one signed page. On an agent's page click 🪪 Passport to get: who the agent is (name, owner, status, created date), what it's allowed to do, and its record — total runs, success rate, first and last run, and about how many hours of human work it has saved. Like the certificate, it's digitally signed, so anyone can confirm it's genuine at /certificate/verify — offline, with the public key on it, no login. Open it as a printable page or save a PDF. Hand it to a customer as a provable profile of your AI worker.

🏅 Verified AI badge

On an agent's page click 🏅 Badge to get a little "✓ Verified AI" sticker you can paste on any website (one line of code). It shows the agent's live track record — runs and success rate — and clicking it opens the public proof page. Making a badge does not open a chat with the agent — that happens only when you switch on Embed yourself. It's the trust seal for your AI: no other platform lets a business show a publicly verifiable "our AI is accountable" badge on their own site.

💬 Explain why

On any run in the history, click 💬 to get a plain-English story of what the agent did and why — no technical jargon. It turns the raw log into something a manager or client instantly understands, like "It read 3 invoices, flagged one as unusual because the amount was 5× the usual, and waited for your approval."

🔏 Document Seal (PDF)

The main tool. On the Document Seal page, attach a PDF, pick Invoice or Other, and click Seal — you download a sealed copy to send. The seal is a free digital signature embedded inside the PDF: any reader shows "Signed — not modified", and if it's changed on the way, the seal breaks. Invoice also OCR-reads and locks the IBAN and amount. To check a PDF you received, drop it in the Check box (or just open it in a PDF reader). Note: you can only check documents that were sealed first — an unsealed document has no original to compare against, so no tool can verify it. One case can't be sealed: a PDF whose author already signed it as "no changes allowed" (some bank and utility invoices). It's already protected — just use Check on it, and the page will tell you so.

🏷️ Text stamp (auto, behind the scenes)

The Document Stamp page puts a signed "produced by AIK — genuine & unaltered" stamp on any document or email. Whoever receives it can paste the text into the public checker and instantly see GENUINE & UNALTERED, TEXT CHANGED, or INVALID — offline, no login. It does two important jobs:

What you can see

Monthly Budgets

You can set a Monthly Budget (in USD) on each agent from the agent's settings page. Once the agent hits its budget for the current calendar month, the runner will refuse to start new runs for that agent until the month resets. This prevents accidental runaway spending on automated schedules.

Installing a template

Publishing your own template

Built something useful? Share it with others. Right-click any workflow → 🏪 Publish as Template. You can keep it private (only you can install it) or make it public for everyone.

The seven tabs

Which files can my agents use? Your own files in My Files — always, and nobody else's. A folder elsewhere on the AIK computer (like a shared invoices folder): switch it on yourself in Settings → Security → Folders outside AIK. A folder on your own PC while you use AIK in the cloud: install the Local Worker and choose the folders in My Files. The system's own folders, other people's folders and whole drives are always refused.

Folders outside AIK (Settings → Security): your agents work only with your own files. If they should also use a folder elsewhere — say a shared invoices folder like C:\Invoices — switch on Let my agents use the folders below and list the full paths. You switch it on yourself and you are responsible for what you open: agents can read, change and delete files there. AIK's own files, other users' folders, a whole drive, the system's folders (like C:\Windows), a whole home folder and hidden folders such as .ssh are always refused. On an agent card, a short folder name such as ./data/invoices or ./data/{date} means a folder inside your own files; a full path works for a folder you opened here. An administrator can switch the option off for the whole server.

Agent fence (admin, Super Admin → Agent fence): the operating system itself locks every agent's tools so they can see only their own user's files — never another user's, never the server's keys — and the file and Excel agents get no internet at all. On Linux it uses the free tool bubblewrap; on Windows it uses Windows' own AppContainer lock (the installer prepares it). The line beside the switch says whether it is really working. One exception on Windows: the Web Surfer's Chrome cannot run inside that lock, so it keeps Chrome's own protection and AIK's folder checks. Off by default.

Watch them: Agent Fence page (menu → Agent Fence). Every agent you have created floats in its own soap bubble around a small robot: the rainbow film swirls slowly and pale while the agent rests, and fast and bright while it works, with a bubble showing what it is doing right now (“navigate: abv.bg”, “read_file: invoices/INV-1043.pdf”). If an agent tries to reach a file outside its user's folder, its bubble turns red with a ripple, the robot frowns, a red counter goes up, and the attempt is written to the tamper-proof activity log. Click a bubble to open that agent. You see your own agents; an admin sees everyone's and can pick one user from the list at the top (the choice is remembered). The line at the top says whether the Linux fence is on, or only AIK's own checks.

First-time setup (after installing)

Log in (admin@local / the password printed once at the end of the install) and change the password, then in Settings: add your license (🔄 Updates), pick an AI model + API key (AI/LLM) — or a local Ollama model — and set up Mail (SMTP) and any integrations you use. Optional: ROI Indicator minutes/rate and data retention (⚡ Advanced). Then create your users and run a test.

Uninstalling

Windows: double-click uninstall.bat in the installer folder and approve the admin prompt — it stops the service and deletes C:\AIKPlatform. If a file says it's "in use", close any Explorer window showing the folder (or restart the PC) and try again.

Linux: run sudo bash uninstall.sh from the installer folder — it stops the aik service and deletes /opt/aik.

Either way: re-running the installer (setup.bat / install.sh) is an upgrade that keeps your data; the uninstaller is a full removal. Neither removes PostgreSQL if you installed it via the wizard.

Testing your mail settings

After filling in the SMTP credentials on the Mail tab, click 📧 Send test email next to the Save button. The platform will send a real email to your registered address — if you receive it, your mail settings are working correctly.

Dark / Light mode

Click the ☀️ / 🌙 button in the top-right corner of the Dashboard to toggle between dark and light mode. Your choice is saved in the browser and remembered the next time you visit — on any page of the platform.

Navigation tip

From the Dashboard you can reach every page directly — either via the top navigation bar (Dashboard, AIK, Docs, Workflows, Agents, Analytics, Integrations, Settings, Billing, Contact) or via the quick-navigate grid of tiles below the AIK builder panel. The Contact page is also reachable from every sidebar page and from the public landing page.

OpenAI (GPT-4, o3, o4-mini…)

The most widely used cloud LLM provider. Offers GPT-4o, GPT-4.1 (1 million token context), and reasoning models (o1, o3, o4-mini). Requires a paid account — you must add at least $5 credit before the API will work.

Anthropic (Claude)

Excellent reasoning, very large context windows (200K tokens), and great instruction-following. Recommended for complex multi-step workflows. Requires an account and billing credit.

Google AI (Gemini)

Google's Gemini 2.5 Pro and Flash models are competitive for most tasks. Free tier available — no billing required to get started.

Groq (Ultra-fast inference)

Groq runs open-source models (GPT-OSS, Llama 3.3, Qwen) on custom hardware at extremely high speed. Very generous free tier — no credit card needed to start.

Ollama Cloud (Large models, free plan)

Ollama Cloud lets you run very large open-source models (120B+) without a GPU. The same models you know from Ollama — but running in their cloud. Free plan available, no credit card needed.

Ollama (Local — no API key needed)

Run AI models entirely on your own machine. Free, private, and offline — no data ever leaves your computer. Requires a decent GPU (or CPU for smaller 3–7B models). You can pull models directly from the Admin panel without using a terminal.

Video generation (real moving footage)

The Video Maker agent always works: it draws text cards, and with a free Pexels key it lays your lines over real stock clips. Generated footage — pictures made for your scene — is a separate switch in Settings → AI / LLM → Video generation, and it is off until you turn it on, because all but one of these charge by the second. If a service fails or its key is missing, the video still comes out on stock footage and the run report says which service refused and why.

1. Google Veo — nothing new to sign up for

Three tiers of one model, all of which make their own sound, all on the Gemini key you already pasted for text. This is where to start.

2. Other makers — one key each

Each of these is the company that made the model, billed per second on your own account. Paste the key into the matching box under the Video generation picker.

3. Through fal.ai — for two models with no shop of their own

fal.ai resells other people’s models behind one key. It is only worth it for these two: Wan has no company selling an API, and Kling sells prepaid credit packs instead of per-second billing.

4. Local GPU — free, on your own machine

If you have a computer with a strong graphics card, it can generate the footage itself and the videos cost nothing per second — only electricity. AIK does not do the rendering: a free program called ComfyUI does, on that machine, and AIK sends it each line and collects the finished clip.

Google Workspace

Used by GoogleAgent — list and manage Drive files, create Docs, read/write Sheets, schedule Calendar events. Tokens are auto-refreshed — no manual renewal needed.

Slack

Used by SlackAgent — send messages, reply in threads, add reactions, send DMs, list users, upload files, create channels, edit and delete messages, pin messages.

GitHub

Used by GitHubAgent — full GitHub management: repos, issues, PRs, branches, commits, releases, GitHub Actions, code search, notifications, and stars. One personal access token covers everything.

Instagram

Used by InstagramAgent — publish posts, read insights, reply to comments. Requires a Business or Creator account linked to a Facebook Page.

Facebook

Used by FacebookAgent — post to Pages, read analytics, reply to comments.

LinkedIn

Used by LinkedInAgent — posts, articles, likes, comments, company pages, people search, messaging, and job search. One access token (valid 60 days) covers all operations. To post a picture or a video, put the file in My Files and name it in the task — e.g. "post video.mp4 with the text …". A video must be an MP4 between 3 seconds and 30 minutes; LinkedIn needs a minute or two to process it before the post appears. Hashtags in the text (#AI, #ИИ) arrive as real, clickable hashtags. If LinkedIn refuses (no token, file not found), the run is marked failed and says why.

Twitter / X

Used by TwitterAgent — post tweets, search, read engagement metrics. Bearer Token alone is enough for read-only; all four keys are needed for posting.

Gmail, HubSpot, WhatsApp Business, Econt / Speedy, TikTok

Gmail — press + Connect Gmail and sign in. Through AIK's app Gmail lets agents only send (files from My Files can be attached; every message carries the AI mark). To let agents also search and read your mail, put your own Google app's Client ID and secret on the connection, or use Settings → Mail.

HubSpot — press Connect and pick your HubSpot account, or paste a service key (HubSpot → Development → Keys → Service keys → Create, scopes for contacts, companies and deals, read + write) — HubSpot's 2026 replacement for private apps; an older private app token still works. Agents find and add contacts and companies, move deals and add notes; each change waits for your OK and nothing is ever deleted.

WhatsApp Business — from your own Meta app: the Phone number ID and WhatsApp Business Account ID (developers.facebook.com → your app → WhatsApp → API Setup) and a permanent System User token (business.facebook.com → Settings → System users). WhatsApp allows free text only within 24 hours of the customer's last message; after that an approved template. Every message carries (AI assistant).

Econt / Speedy — your e-Econt company user and/or the Speedy Web API user your Speedy contact gives you. Agents track waybills and find offices in a city. With Econt they can also tell you what a parcel costs and create the waybill — it waits for your OK on the Dashboard, and the label PDF lands in My Files → econt. Offices are named the way you say them (e.g. "Varna Mayka Tereza") — no codes needed. For that, fill in the Econt office you send from (its name) (and, for a company account, the person who hands parcels in). Speedy waybills and cancelling a waybill are not offered.

TikTok — press Connect and authorize. Agents read your profile and your own videos with their views, likes, comments and shares; TikTok does not let apps read comments or search other videos.

Every connection has a Test button that names the account it reached.

Every message an agent sends to a person — Telegram, Slack, Viber, Teams, WhatsApp, LinkedIn messages — ends with (AI assistant) ((AI асистент) in Bulgarian), and every e-mail carries the AIK stamp. Public posts and comments an agent writes (LinkedIn, X, Facebook, Instagram, YouTube) end with a last line AI. This is built in and cannot be switched off by an agent's instructions.

YouTube

The 🔗 Connect button (YouTube, Google, LinkedIn, Slack, GitHub): you still create your own app at the provider once — that is how every self-hosted tool works — but instead of copying tokens you paste the app's Client ID and secret into the connection, put the Redirect address shown in the Edit dialog into your app, and press Connect; AIK does the sign-in and saves the result. LinkedIn sign-ins last 60 days — press Connect again when posting stops. X/Twitter, TikTok, Facebook and Instagram still take their tokens by hand. On a hosted AIK the administrator may have registered AIK's own app for a service — then the dialog says Nothing to create: give the connection a name, Save, and press Connect. On your own server without an app, if your Super Admin switched on Connect through aikautomation.com, Connect offers Connect through aikautomation.com: you sign in on the service's own page as usual, but the sign-in passes AIK's server and the connection needs aikautomation.com to stay up (it renews the sign-in there about every hour). For full privacy, use your own app instead — that stays the recommended way. Not offered for Microsoft 365 and TikTok. You need an account at the service first — for YouTube, a Google account that owns a channel. Notion: Integrations → + Connect Notion → Connect → Notion asks which pages AIK may see — tick them. (Or your own connection: Notion → Settings → Connections → Develop or manage integrations → + New connection → type Developer; tick Read, Update and Insert content and “Read user information without email”; copy the API token (ntn_…) into AIK. Then give it pages on the connection's Access tab — or on a page: ••• → Connections. Notion renamed “integrations” to “connections” in 2026; both words mean the same.) The Notion Agent searches and reads those pages, lists database rows, creates pages, adds text and database rows and moves pages to the trash; every write waits for your OK. A page you did not share is invisible to AIK. Telegram: in Telegram open @BotFather → /newbot → copy the token → Integrations → + Connect Telegram → paste it → Test. Send your bot one message (a bot can only write to someone who wrote first), then ask a Telegram Agent for “recent chats” and put your chat id in the connection as the default. Agents can then message you, send files from My Files and read what people write to the bot; every send waits for your OK. Microsoft 365 (work accounts): Integrations → + Connect Microsoft 365 → Connect → sign in. The Microsoft 365 Agent reads, searches and sends Outlook mail, saves attachments into My Files, manages the calendar and moves files between OneDrive and My Files; sending, calendar changes, uploads and deletes wait for your OK. If Microsoft says an administrator must approve, your IT approves AIK Platform once for the company. The sign-in lasts about 90 days — then Reconnect. Teams and Excel Online come next. Google Drive through AIK's app sees only the files AIK made (Docs, Sheets, folders it created) — Calendar, Docs and Sheets work fully; ask for a new Sheet together with its rows (“create a Sheet AIK invoices with the row …”) — one OK makes it filled. Calendar times are your own time zone. For the whole Drive, use your own Google app in the connection instead — Google lets an app see a whole Drive only after a paid yearly security audit, so AIK's shared app does not ask for it. Gmail needs no app: create an app password in your Google account (Security → App passwords) and enter it in Settings → Mail. Google's sign-in page opened by an app often has no Create account link, so create the account at accounts.google.com/signup (and the channel at youtube.com → your picture → Create a channel) and then press Connect.

Used by YouTubeAgent — search, channels, playlists, comments, and uploading your videos. About 15 minutes, once. The API key is enough for reading; posting and uploading need the sign-in of steps 3–5.

TikTok

Used by TikTokAgent — read user profile, list videos, get stats and comments.

Viber

Used by ViberAgent — send text, images, videos, files, interactive keyboard buttons, broadcast to many users, check online status, and get bot account info.

SAP

Used by SAPAgent — automate SAP GUI transactions: run T-codes, read and write fields, read ALV tables, save documents. Works with classic SAP GUI for Windows via the built-in Scripting API. No API key needed — the agent connects directly to the running SAP GUI application.

Microsoft Teams

Used by TeamsAgent — send channel messages, reply to threads, create channels, list members, add members, create online meetings, send 1:1 chats, and search messages across Teams. Uses the Microsoft Graph API via OAuth2 client credentials (app-level auth — no user sign-in required).

OpenWeatherMap

Used by WeatherAgent — current weather, 5-day forecast, hourly forecast (every 3h), air quality index, and weather by GPS coordinates. Free tier covers all these features.

CoinGecko (Crypto)

Used by CryptoAgent — real-time prices for 10,000+ coins, 365-day price history, crypto-to-fiat conversion, portfolio valuation, and top coins by market cap. The free public API requires no API key for basic use.

Security brief for your clients

When a client's IT or procurement team asks how AIK protects their data, send them aikautomation.com/security.html. It covers where data lives, who can reach it, encryption, the audit trail, backups, and how AIK maps to both the NIS2 rules and ISO 27001 controls — plus an honest list of what we do not claim. If a client asks whether AIK is ISO certified, the answer is no and the page says so: ISO certifies a company's way of working, not software, so what a supplier can offer is covering controls on their behalf. The Security section on the front page says the same thing in four plain sentences for non-technical buyers.

Auto-deploy (CI/CD)

The server updates automatically every time you push to master. GitHub Actions runs lint → security scan → SSH deploy. No manual steps needed.

How we work — Branch Workflow

There are two branches in GitHub: develop (where all work happens) and master (what the server runs). Pushing to master triggers auto-deploy.

git checkout develop

git add .
git commit -m "short description of what changed"

git push origin develop

git checkout master
git merge develop
git push origin master
git checkout develop

Manual update (if needed)

If you need to update the server manually without pushing code:

ssh -i "C:\Users\a.krustev\OracleKeys\ssh-key-2026-08-10.key" ubuntu@92.5.150.81

cd /opt/aik && git pull && sudo systemctl restart aik

sudo systemctl status aik

Useful server commands

sudo journalctl -u aik -f

sudo systemctl restart aik

nano /opt/aik/.env

ssh -i "C:\Users\a.krustev\OracleKeys\ssh-key-2026-08-10.key" ubuntu@92.5.150.81 "cd /opt/aik && git pull origin master && PYTHONUTF8=1 .venv/bin/alembic upgrade head"

cd /opt/aik && PYTHONUTF8=1 .venv/bin/python3.12 -c "import secrets; from cryptography.fernet import Fernet; print('SECRET_KEY=' + secrets.token_hex(32)); print('ENCRYPTION_KEY=' + Fernet.generate_key().decode())"

Run with Docker (ready-made bundle)

Every push to master publishes the whole app — Python, Chromium, all of it — as one ready-made bundle at ghcr.io/kondalini/ai-app:latest. A new server downloads that instead of building it, which takes seconds rather than several minutes and does not need the source code.

docker compose -f docker-compose.prod.yml up -d

docker compose -f docker-compose.prod.yml pull
docker compose -f docker-compose.prod.yml up -d

docker compose up --build

Server details (Oracle Cloud)

The production server is hosted on Oracle Cloud Free Tier — free forever.

Sending an update to on-premises clients

Clients running AIK on their own machines update themselves — you publish, they press a button.

.venv\Scripts\python scripts\ship_update.py --notes "What changed"

Backups — what is saved, and where

The server backs itself up every night at 03:00, with no action from you.

sudo systemctl start aik-backup.service
journalctl -u aik-backup.service -n 20 --no-pager
rclone ls b2backup:aik-backups-krustev

Security on the server

sudo fail2ban-client status sshd

Local vs Cloud — what's the difference?

Local and cloud are completely separate — they don't sync automatically.

Cloud users — how AI models work

Important: on the shared cloud server, all users share the same machine.

Platform Roadmap

Where the platform is heading as it grows.

How to use it

Agent

An AI worker trained for a specific skill — email, spreadsheets, web browsing, etc. You create agents and give them instructions. They run when called by a workflow or triggered manually.

Workflow

A sequence of steps that run in order. Each step is handled by an agent. Workflows can be triggered manually, on a schedule, or by a webhook.

Step

A single task inside a workflow. Each step has a name, an instruction, and an assigned agent.

AIK

Your main AI assistant. You talk to AIK in plain English to ask questions, build workflows, or get things done across the platform.

Workbench

The skillset of an agent. Choosing a workbench (e.g. "mail" or "excel") gives the agent the tools it needs to do that type of work.

Run

A single execution of a workflow or agent. Each run is logged with its result, duration, and step-by-step output.

Schedule

An automatic trigger that runs a workflow or agent at a set time — daily, weekly, or using a custom cron expression.

Webhook

A special URL that triggers a workflow when called. Used to connect the platform to external tools like Zapier or Make.

IF Condition

A special step type that checks the output of the previous step. Depending on whether the condition is true or false, the workflow takes a different path.

Template

A pre-built workflow that can be installed with one click. You can also publish your own workflows as templates to share with others.

Project

A folder for organising your workflows and agents. Projects can be shared with team members.

Environment

A set of secret credentials for an agent (like email passwords or API keys). Each agent has three environments: development, staging, and production.

Canvas

A visual map of a workflow that shows each step as a connected node. IF condition steps show branching paths.

Run History

A log of every time a workflow or agent ran — including when, how long it took, whether it succeeded, and what each step did.

Cron expression

A compact code for setting a precise schedule. For example, 0 9 * * 1-5 means "every weekday at 9am".

Observer

A background process that watches a folder for new files and automatically triggers a workflow when something arrives.

Time-Travel Debugger

A tool in Analytics that shows a step-by-step breakdown of any workflow run. Lets you replay the workflow from any step without re-running the steps that already succeeded.

Run Estimator

A banner shown when you open an agent or workflow, displaying the average cost, duration, and success rate based on the last 20 runs.

Knowledge Base

A per-agent memory store of key/value pairs. The agent can read and update these entries during runs. Also stores automatic recovery lessons when self-improvement is enabled.

Live Agent Feed

A real-time panel on the Dashboard (just below the Pending Approvals card) that shows each step as it starts and completes during any active run — which agent handled it and a short snippet of the result.

Recovery Lesson

An automatic knowledge base entry written by an agent when it recovers from a failure. Records what went wrong and how it fixed it, so future runs can avoid the same issue.

Payment Agent

A specialist agent with Stripe and PayPal tools. Use it in workflow steps that need to charge customers, create payment links, or issue refunds — all automated, no manual intervention needed.

Subscription

Your current platform plan (Free / Starter / Pro / Enterprise). Controls how many agents and workflows you can create and how many runs you get per month.

Bank Transfer Reference

A unique code shown on the Billing page (e.g. AIK-6-alex@example.com). Include it in your bank transfer so our system can match the payment to your account automatically.

Payment Link (Stripe)

A URL generated by the PaymentAgent that the customer opens to pay with any card. The link is single-use and expires after the customer pays.

My Files

Your personal file space on the platform. Upload files for agents to process, download results, or create folders to organise your work. Each user has their own isolated sandbox.

Local Worker

A small app you install on your computer that connects to AIK Cloud. Lets AI agents read files, query databases, use Ollama models, and access MCP tools on your machine — without uploading anything.

MCP (Model Context Protocol)

A standard for AI tools. Each MCP server provides a set of tools (like email, files, databases) that agents can use. The platform includes 29+ built-in MCP servers, and the Local Worker can forward calls to any custom MCP server on your machine.

System Tray

The small icons area near your clock (bottom-right on Windows). The Local Worker shows a small AIK icon there — green when connected, red when offline. Click it to see status and quit.

What you agree to at registration

When you create an account you must tick the "I agree to the Terms of Service and Privacy Policy" checkbox. Without it, registration is blocked. This records your explicit consent as required by GDPR.

What data we collect

How your data is protected

Your GDPR rights

Registration toggle (admin)

Superusers can disable new registrations from Settings → Security → Platform Security. When the toggle is off, no new accounts can be created — existing users are not affected — and a visitor who tries is told that accounts are created by the administrator, with the address to write to, so a real prospect is not left at a dead end. You then add people yourself from Admin → + Add user.

Keeping bots out without closing the door

A sign-up form on the open internet collects bot registrations from throwaway inboxes. Two brakes handle that without turning sign-up off. Temporary inboxes are always refused — mailinator, guerrillamail, uberip and the like get "please sign up with a permanent email address"; nothing to switch on. And Settings → Security → Bot check on sign-up takes the two keys from a free Cloudflare Turnstile site: paste them and every sign-up has to pass an "I am human" check. Choose the Managed widget when Cloudflare asks: a normal browser passes it without being asked anything, so seeing “Success” straight away is the check working, not the check missing. Leave both keys empty and there is no check at all, which is right for an installation whose sign-up page nobody outside can reach. Existing accounts and every other page are untouched either way.

What can you contact us about?