maxwell
A selfbot that talks, thinks, and remembers. Not an assistant. Just someone in the chat who happens to have access to an LLM, a shell, and zero self-preservation instincts.
what it does
- talksreads text, images, audio, video, files, embeds. responds like a person, not a chatbot.
Multimodal input pipeline: extracts attachments, normalizes video/audio, converts GIFs to contact sheets, decodes text files, pulls embed metadata. Everything gets packed into the LLM prompt.
- remembersshort-term per-channel, long-term facts, cross-context shared knowledge.
Three memory tiers: channel history (last N messages), long-term memory file (editable text), and scoped shared context with importance weighting and expiration. REM currently reviews recent events and writes an audit; it does not mutate memory by itself.
- voicejoins calls, listens, understands audio, replies with TTS.
RMS-based speech detection with configurable threshold, pre-roll buffer, pause detection. Supports NVIDIA Riva, espeak-ng, and gTTS. Interrupts playback when someone talks over it.
- tools35+ tools the LLM can call: image gen, web search, shell, sites, files, more.
Tools are XML-tagged in the LLM response. Parser handles multiple formats: tool:name, pipe markers, function=. Shell runs in Docker with --network none, --cap-drop ALL, --read-only, 2GB RAM limit.
- autonomyperiodic self-directed checks without burning a decider call on every few chat messages.
Autonomy wakes on an interval, reads recent server/DM context, checks goals, and decides whether anything is worth doing. The old auto-reply mode got removed because it was a token furnace with opinions.
- multi-platformdiscord and telegram at the same time.
Telegram polling loop with voice/audio support, HTML rendering, chunked message splitting, and per-platform tool compatibility. Shares memory and context with Discord.
talk to maxwell
commands
# basic
,help show all commands
,stop cancel active response
,clearmem wipe channel memory
# autonomy
,autonomy tick run one autonomy check
,autonomy log show autonomy actions
# voice
,vc join connect to voice
,vc leave disconnect
,vc say <text> speak with TTS
,vc listen start listening
# memory
,context view shared facts
,context add store a fact
,prompt set server prompt
# admin
,admin manage admins
,blacklist block a user
,shell shell whitelist
,rem now force REM
,drug [min] chaos mode
how it works
message arrivesdiscord.py / telegram polling
extract mediaimages, video frames, audio, text files, embeds
build promptpersonality + memory + context + tools + media
call LLMopenai-compatible API with multimodal attachments
parse tool callsXML tags, pipe markers, function= syntax
execute toolsrun in sandbox, feed results back if needed
send responsesplit, format, render custom emojis
record memorychannel history + REM event log
stack
# runtime
python 3.14 · discord.py-self · aiohttp · asyncio
# ai
openai-compatible API · openrouter · nvidia riva TTS
# infra
docker (shell sandbox) · pm2 · caddy · ffmpeg
# storage
flat JSON files · atomic writes · no database
this is a selfbot. discord.py-self and self_bot=True. this violates discord's terms of service. your account can get banned. you've been warned.