Superpowers for your
Claude Code agent

Chat with a headless Claude Code agent over Telegram. Every message runs a real Claude turn with full tools; each chat keeps its own persistent conversation. Send files — the bot ingests, analyzes, edits (posted PDFs included) or holds them, and common "send me that file" requests are answered sub-second by deterministic reflexes, no LLM turn at all. Allowlist-locked, works behind NAT.

Talk to your own agent by voice from the Agent Voice Mode iOS app. A tiny stdlib webhook relays each spoken question to your agent CLI (Claude Code out of the box) and speaks the answer back — pair by scanning a terminal QR. No API keys, no voice code on your machine: speech runs hosted-side.

Label your data public/private and route privacy-sensitive chat queries to a private local LLM with full context — the cloud agent never sees the underlying records. Fails closed, with a /cloud escape hatch.

A grammar, spelling and clarity fixer that runs entirely against a local model — your text never touches the cloud. Fails safe: if the local model is down it blocks rather than leaks.

Turn a Telegram group into a self-filing lab notebook for one R&D project: every post — text, voice notes, photos, documents — is auto-filed, annotated and semantically indexed before the chat turn runs, and questions are answered from the project's own files first.

Turn a Gmail mailbox into a realtime chat channel for an AI agent. IMAP IDLE sees new mail in seconds; every verified email from the owner runs an agent turn and gets a threaded reply. Mail from anyone else never triggers the agent — it's reported to Telegram instead.

Auto-drafts replies to inbound product inquiries in the owner's own voice — grounded in a product KB, threaded correctly — and learns from the replies you actually send, so drafts get closer over time. Optional PII-masking layer before anything reaches the cloud.

Event-driven new-mail watcher: a persistent IMAP IDLE connection sees new mail within seconds — no polling — and drops each qualifying message as a JSON job for a downstream consumer, like a chat gateway that runs the email as a chat turn.

A cron pipeline that turns your mail archive into a living knowledge base + lightweight CRM: contacts, durable product knowledge, order/quote facts, rolling per-contact summaries with inline source citations — and it learns your writing style too.

A local file-plus-SQLite CRM: one human-editable markdown file per contact beside a SQLite archive of all sent/received mail, with deduplication, alias folding and a rolling activity summary. The instant answer to “who is this and where were we?”

A daily digest of customer inquiries still awaiting a reply. Scans all connected mailboxes, finds threads where the customer spoke last, asks an LLM whether each really needs an answer, and emails you a clean actionable list every morning.

Read, search, draft and send across any number of Gmail mailboxes from one small CLI over OAuth — each account with its own token, selected per command. Saves drafts by default; sends only when explicitly told.

Every reply you send is a free training example. This loop tests whether your knowledge base alone could reproduce your real answers — in facts and in style — then patches the gaps automatically. Your KB gets smarter with every email.

Dependency-light semantic search over a folder of markdown knowledge, plus a grounded tier-1 quick answer: cheap local retrieval handles the easy questions, and only genuinely hard ones escalate to the big model.

Private document ingest + retrieval-augmented Q&A: point it at PDFs, CSVs and text; it parses, chunks, embeds and indexes locally, then answers with citations to file and page. With a local backend your documents never leave the machine.

Neural image & video search over a local media library: one text query is matched in CLIP space against both what an image looks like and what you said about it — the better score wins. Sub-second results.

A searchable SQLite/FTS5 log of every chat message and assistant reply, with automatic LLM-driven project tagging and multi-label classification. Recall any past thread by keyword, date, project — or meaning.

Read the real, JavaScript-rendered content of any page — including SPAs whose HTML is an empty shell until scripts run. Renders the DOM in headless Chrome, then parses it to clean text or structured data. Now also fetches images from sites whose WAF blocks plain downloaders, by capturing them straight from the browser's own network layer.

Dependency-free shell templates for reliably starting long-running services on boot and keeping them alive — covering the three service shapes every agent box ends up running.

Two scheduled self-maintenance routines: a weekly local health check that grooms memory and audits crons, processes and log bloat, plus an auth watchdog — alerts arrive on Telegram only when a judgment call is needed.

Scheduled compressed backups of a project folder to Google Drive, keeping only the newest N copies. Regenerable bulk and secrets are excluded, so archives stay small and safe to store.

Auto-generates competitive-research reports on a schedule, renders each to a styled A4 PDF and distributes them by email, chat and cloud drive. One engine, any number of topics — you just add a prompt file.

Business letters on company letterhead with the logo pinned to a fixed spot on the page — it won't drift or reflow across Word, LibreOffice, Google Docs or PDF export. Template pack + the anchoring technique explained.