🌍 Languages: English | Deutsch

---

AIfred Intelligence

A self-hosted AI assistant that gets things done — tool chains, sub-agents, multi-agent debates, voice, camera surveillance and local LLM inference on your own hardware.

AIfred runs on your own machine and works for you across every channel you use: it reads and answers e-mail, books appointments, manages contacts, documents and databases, watches your cameras and talks to you through the browser, Telegram, Discord or a voice terminal in your living room. Local models, persistent memory, no cloud dependency — your data stays with you.

📺 Example showcases — exported chats: multi-agent debates, chemistry, maths, coding and web research.

⭐ If AIfred is useful to you, please star the repo. Self-hosters easily forget — but stars are the main signal that tells me people are actually using this. It directly affects whether I keep building.

At a glance

• 🔗 Autonomous tool chains across plugins — one prompt, many steps, from any channel
• 📬 Your mail, read properly — unread mail, several at once, PDFs and Office files as text, scans and photos through the vision model; nothing is cut off, long mails are read in steps
• 📅 Nextcloud as your hub — calendar, tasks, contacts, notes and files, with rights per module
• 🌅 Morning briefing — by mail every morning: the next seven days of appointments, birthdays, open tasks and unread mail, anything with a deadline first
• 🎩 Multi-agent system — AIfred, Sokrates, Salomo and your own agents debate, criticise, judge — or take delegated tasks as sub-agents
• 📡 Reachable everywhere — web UI, Telegram, Discord, e-mail, Echo Dot voice terminal; runs headless
• 🎤 Voice — speech-to-text plus the TTS-Suite service over seven engines (local, other hosts, cloud, browser), voice cloning, spoken announcements on the Echo Dot
• 👁️ Vision & Vigilantia — image analysis in chat and continuous camera surveillance with face recognition
• ⚙️ Local inference, tuned automatically — llama.cpp, vLLM and Ollama with VRAM-aware context calibration across multiple GPUs
• 🔒 Security built into the framework — permission tiers per channel, prompt-injection defences, credential broker, audit log

---

🔗 Complex Workflows (Tool Chains)

The real value doesn't come from individual tools but from chains that AIfred works through on its own across several plugins — from any channel: browser, voice terminal (FreeEcho.2), Telegram, Discord, e-mail.

Example (real, tested end to end): you photograph a business card with your phone and say — typed or spoken:

"Analyze this business card, extract all the data, create the contact in Google, check next week for a free morning, book a two-hour appointment, and send me the photo along with the contact details via Telegram."

AIfred runs this as one continuous chain:

1. 👁️ Vision — the VLM reads the card (name, phone, e-mail, address, web) directly from the image
2. 📇 Google Contacts — creates the contact in structured form (google_contacts_create)
3. 📅 Google Calendar — first checks the time window for free slots (google_calendar_list_events) and only then books the appointment; an existing appointment is recognised and AIfred asks back instead of creating a duplicate
4. 📤 Telegram with attachment — sends you the photo of the card plus the contact details — without knowing your chat ID (owner default)

All local, a single prompt, no clicks in between. Files from the conversation (uploaded images, PDFs or plots from the sandbox) can be sent as attachments over any channel, isolated per session.

Second example (real, tested end to end): mail with a scanned invoice and a PDF report attached:

"Which mails are unread? One of them has an invoice as a scan and a report of a test drive. Read both attachments and give me only the invoice number, the amount, the due date and the driver's name."

1. 📬 Unread mail — lists only what you have not read yet; reading changes nothing in your mailbox
2. 📖 Several at once — reads the mails in a single call, HTML turned into plain text
3. 📄 PDF — reads the report through the same parsers as the document index
4. 👁️ Scan — the invoice has no text layer, so its pages go to the vision model, which ran beside the chat model without unloading it

No fixed limit decides what AIfred may see: how much fits in one step follows the model's context, and long mails and attachments are read by lines, like files. Text from a mail stays marked as foreign, so a mail cannot give AIfred orders.

Prerequisites & knobs:

• Security tier: creating things (contact, appointment, running code) needs WRITE_DATA or higher. In the browser you always have it; external channels are raised as needed in the Plugin Manager — deliberately only there, no sender can grant itself rights. See Security
• Vision + tool chains: for reliable tool calls on the image path, turn thinking off for the vision LLM (🧠 icon; the 💭 reasoning prompt may stay on). Reasoning models with speculative decoding (MoE/MTP) otherwise swallow tool calls after thinking

---

✨ Features

🧠 Autonomous Agent & Tools

The LLM decides on its own which tools to use — OpenAI-compatible function calling, up to 100 tool calls per request, every tool delivered by a plugin:

This is the beginning of the README. Read the rest at the original. Read the full README ↗