No plugins to install, no cloud subscriptions, no third-party connectors. Everything runs locally in one container.
Each agent runs its own Telegram bot. Add several to one group — they take turns, never loop, only reply when addressed. Each has isolated context.
Send voice notes in Telegram. humux transcribes with faster-whisper and responds by voice via edge-tts or Kokoro.
Approve or reject write actions directly from Telegram inline buttons. Permission rules use glob patterns: ALWAYS, ASK, NEVER.
Acknowledge messages with emoji reactions instead of text — thumbs-up, heart, eyes. Pre-approved, no prompt interruption.
Gate who can trigger an agent in each group and who may DM it: everyone, nobody, or specific user IDs.
One turn can produce several heterogeneous messages — text, voice, images — delivered as separate bubbles via a [[split]] marker.
Read and send WhatsApp messages via the wacli CLI tool. Integrated as a skill the agent can learn.
Chat with the agent directly from your terminal — SSH in or run uv run python -m core.cli. Resumable sessions, no Telegram needed.
Read, compose, search, and manage emails. Powered by Himalaya CLI, a battle-tested Rust MUA.
Each agent can own a dedicated mailbox or be granted read / read-write access to your inbox. Bind accounts per agent.
Credentials resolve from the encrypted vault — they never enter the model's context. Bitwarden import supported.
Check your schedule, create events, update existing ones. CalDAV works with Google Calendar, iCloud, Fastmail, and any CalDAV server.
Bind calendar accounts per agent with read or read-write permissions. Granular control over who sees and modifies what.
Morning briefings include today's events, emails, and reminders. Cron-based scheduling for any custom task.
Search contacts, create new entries. Supports CardDAV (Purelymail, iCloud, Fastmail) and Google Contacts via the contacts CLI.
Bind contact books per agent with read or read-write access. Agents only see what you grant them.
Register a GitHub App and receive issue, PR, and comment events in near-realtime. Each event triggers an agent turn with full context.
Route events to multiple agents based on @mention slugs and author filters. Each agent gets its own isolated conversation thread per event.
Visual feedback on webhook processing: 👀 when picked up, 🚀 on completion, 😕 on failure. Mid-turn steering injects follow-up messages into running turns.
Create and switch between agent identities. Each has its own character, skill/tool scope, voice, and account bindings. Enable or disable per agent.
Lexical retrieval, semantic search, importance-based forgetting, and automatic deduplication. Extracted from every conversation.
Delegate scoped subtasks to a sub-loop. Run sync (result in-turn) or background (distilled summary). Set step/token/thinking budgets.
Teach the agent new capabilities by writing markdown files — no code required. Install community skills from git repos via the admin UI.
Send a follow-up message while the agent is still working. It gets injected between tool rounds, so you can redirect without waiting for the turn to finish.
Cron-based scheduling for morning briefings, email checks, contact sync, and custom tasks. Pause and resume jobs. Filter by agent in the admin UI.
See token count and percentage of context window used in the Inspect tab. Manage context budget per agent.
Step-by-step first-boot configuration. Walks through API keys, Telegram bot setup, account connections, and agent creation.
Configure agent character, skills, tools, voice, Telegram bot, and account bindings. Tabbed interface with sys prompt, inference, and provider settings.
Track token consumption across 24h, 7d, and 30d windows. Per-provider breakdown, per-agent attribution, and cost estimates in the Overview tab.
Install the admin UI on your phone or desktop. Full PWA support with app icon, standalone mode, and offline shell.
Per-agent log streams filterable by stream, level, time, and text. Inspect context, token usage, and chain-of-thought.
Settings, Long-term, and Short-term sub-tabs with pagination and search. Inline editing via detail modal. Mobile-friendly truncation.
View, create, pause, resume, and cancel scheduled jobs. Monitor subagent runs. Filter jobs by agent.
Encrypted two-tier vault: infrastructure keys (machine-key sealed) and agent secrets (admin-password sealed). Bitwarden import.
Local via faster-whisper or remote via any OpenAI-compatible STT endpoint. Toggle between local and remote in the admin UI. Run Whisper as a sidecar to keep the main image slim.
edge-tts for cloud-quality voices, Kokoro 82M for fully offline multilingual TTS, or any OpenAI-compatible remote endpoint. Switch between local and remote per-engine.
Run voice models in the main container or as Docker sidecars. Whisper and Kokoro have dedicated sidecar services — keep your core image under 1 GB while retaining full voice capabilities.
Tavily or self-hosted SearXNG for real-time web search. Look up current information, news, and research topics — your choice of backend.
Optional headless Playwright browser. Read JS-heavy pages, act on websites with persistent logged-in profiles. Per-domain approval.
Multi-step autonomous research pipeline: plan queries, fan-out web searches, fetch and read sources, verify claims, synthesize a cited report delivered as a shareable artifact.
The agent publishes pages and sites under /artifacts/ using workspace file tools. Sandbox CSP for security. Shareable links.
Two-tier encryption: infrastructure keys sealed with machine key, agent secrets sealed with admin password. Values never enter model context.
Glob-pattern rules: ALWAYS allow, ASK for approval, NEVER allow. Interactive approval via Telegram inline buttons for write actions.
File tools confined to workspace directory. Read pre-approved, writes ask first. Path traversal blocked. Per-call command approval.
Choose from Anthropic Claude, OpenAI, Grok (xAI), Google Gemini, DeepSeek, or OpenRouter. Configure per-agent or globally. Thinking levels up to 'max'. Image generation via OpenRouter, fal.ai, or OpenAI.
Admin UI splits LLM config into Inference, Providers, and History sub-tabs. URL-based navigation for copy-pasteable links.