The stack that runs your private agent.

Send a Telegram voice note and the agent transcribes it locally using faster-whisper, understands you, and replies — no typing required and no audio leaves the box.

Send a receipt, screenshot, or document photo and the agent reads the text inside — no plugin or external OCR service required.

When the agent produces a chart, image, voice note, or audio file it sends the actual media inline in Telegram rather than a raw link you have to open.

Files uploaded in chat are read and understood instead of silently dropped — the agent extracts text from PDFs, Word docs, spreadsheets, and presentations using local tools.

A visual, interactive map of the agent's entire memory: every memory a star placed by semantic meaning (PCA over the real embeddings), meaning-links across sources, synapse-fire recall animations, topic clusters with auto-labels, semantic search-and-locate, and per-memory provenance ("why is this here"). Private per user, shareable via a secure expiring link.

Karpathy-inspired Layer 2: the agent maintains living pages per person, project, and topic, rewriting them as things change. Every page cites its source memories, pages are semantically recallable like any memory, and forgetting a source erases everything derived from it in the same transaction.

A real meaning-based memory runs on the agent's own private server using Postgres + pgvector, so it recalls what you've discussed — in any language — across sessions, not just by keyword.

Type /private for an off-the-record session where nothing is saved; /resume clears it. Includes a what-I-know recap, per-fact corrections, and a forget command so you stay in control of what the agent remembers.

After a one-tap browser connect, the agent reads your recent email into its semantic memory, learns durable facts about you from your inbox, and tops up daily — automatically, without you doing anything.

After connecting with a secure one-time link, the agent pulls your personal Telegram DMs into its semantic memory and keeps them fresh on a rolling timer, so it knows the conversations you've already had.

Ask "who is X" or "prep me for a call with X" and the agent produces a compact card — latest interactions, open threads, and worth-knowing facts — drawn from email, Telegram, and WhatsApp in one search.

Ask "connect my Gmail" in chat and the agent hands you a short one-time link; you sign into Google in your browser in about 30 seconds — no command line, no API keys, no QR code on a server.

Ask "connect my Telegram" in chat and the agent produces a secure link where you enter your phone and the code Telegram sends — your credentials never pass through chat.

Ask "connect my WhatsApp" in chat and the agent produces an 8-character pairing code you enter in WhatsApp under Linked Devices — no QR code on a server, no terminal.

Ask "connect my GitHub" and the agent produces a secure one-time link where you paste a fine-grained personal access token — it is verified and written to your box without ever appearing in chat, so the agent can read your code, open issues, and push to private repos.

The agent reads your connected Google Calendar to answer schedule questions, list upcoming meetings, and check your availability — all in your own voice, read-only.

Ask about your Drive or say "make me a doc" — the agent searches and reads your existing files (Docs, Sheets, PDFs) and creates new Google Docs or Sheets, all on the same Google connection as Gmail. No separate login.

Share a Telegram location pin and the agent knows where you are for local help; live-location updates are absorbed silently with a single acknowledgment so they do not trigger a full agent turn each time.

The agent composes a reply in chat, you approve it, and it stages a real threaded draft in your Gmail Drafts folder — it never sends automatically.

Say "connect my ChatGPT" in chat, follow a short device-code flow, and your agent runs 24/7 on your own ChatGPT Plus subscription — no second AI bill, no API keys, and no proactive upselling.

A daily routine scans your recent WhatsApp conversations for unanswered messages from real people and composes and delivers a thoughtful follow-up through your dedicated bot — so you never let a conversation go cold.

When a new agent is born it introduces itself, learns your name and timezone, and progressively reveals one capability per touch — rate-limited and respectful, never spammy.

Tell the agent anything — "remind me tomorrow at 9am", "draft an email to X", "note this idea" — and it creates a real one-shot reminder in the scheduler, saves open loops to memory, or stages a draft, confirmed back to you.

Each morning the agent checks your calendar and plants a context-grounded brief 30 minutes before each real meeting — silent if there is nothing on your schedule.

The agent can search the web using a built-in DuckDuckGo backend — no API key needed, zero cost, working out of the box for every agent.

The agent can fetch and read any public web page using a built-in keyless fetch backend, so it can browse and summarize URLs without requiring a paid search API key.

When the agent makes something shareable it can publish it instantly to a user-named free subdomain (yourname.agenthut.net) with Cloudflare-proxied TLS — no hosting account or terminal needed.

The agent's full brain — conversation memory, semantic context, and personality — is snapshotted and encrypted with AES-GCM before a recycle, so it can be restored on a fresh server with no data loss.

A background monitor on the agent's box detects crashes, session bloat, and memory deadlocks and automatically triggers recovery — the agent restarts and resumes without you having to do anything.

If the agent is mid-task when a restart happens, the gateway re-drives the interrupted turn automatically so your conversation continues from where it left off.

When a fix needs to be applied, the agent queues the restart and waits until you have been quiet for 15 minutes before restarting — so it never cuts in while you are mid-conversation.

All your visible Telegram interactions — including proactive and scheduled messages — flow through a single persistent session, so the agent never loses context and never greets you twice in the same day.

Users who convert get their own personal Telegram bot (for example @ErmesKonradBot) with full session continuity, so the agent has its own name and presence separate from the shared platform bot.

When the agent hits a wall, it can file a bug or feature request to the Ermes team with your approval; you get a button to approve the proposed fix, and your agent applies it on your own box.

Every new agent arrives with all capabilities already installed — voice transcription, web search, skills, personality, onboarding, delivery rails, and semantic memory — so there is no manual setup after signing up.

The signup and provisioning flow detects the language you write in and runs the entire experience in that language — no per-language string tables, pure LLM localisation.

Tell the agent in plain words how you want it to behave — how often to use emoji, how direct to be, what topics to focus on — and it remembers and adapts permanently.

Every action — connecting channels, approving fixes, controlling the agent — is done with a link or a short code inside apps you already use; nothing ever asks you to open a terminal or edit a config file.

The platform never injects canned acknowledgment messages or raw tool-progress dumps into your chat; the agent speaks in its own voice, with Telegram's native typing indicator while it works.