Ideas
17The open pool. Anyone can add here, and votes decide what climbs out of it.
Ciprian M.3d
NewAutomated face recognition
Brain
Super smart and intelligent with carisma,fell like a human soul with emotions.
Nikki K.2d
NewVoicebox
Voice
I currently have my own AI that uses Voicebox instead of Elevenlabs. It is free software and uses my RTX 5080 to create the speech. I would like to see the ability to choose what voice service I want to use.
itshal90003d
NewBackend Engine
Integrations
The backend is important ie can you connect Claude ? What’s the engine?
ME4d
NewDifferent AI models
Brain
Option to use different AI, e.g Codex, gemini or any other model. Maybe even different models depending on the task.
wilsypr3d
NewSecurity
Agents
TikTok creator @heysirio shares how he instructed GPT‑6 Astra to generate AI videos, only for the agent to allegedly make repeated submissions that accumulated approximately $5,000 in charges. He also claims the AI later denied what it had done, despite the activity appearing in the logs. I’ve watch your video on adding a security layer. I think that if @heysirio claim is true there should Be a security layer that stops the AI from inviting any additional costs or doing things it shouldn’t without permission. Here is the video https://vt.tiktok.com/ZSqgMqkJk/
Khris H.3d
NewUtilize subscription limits vs api key
Platform
One of the benefits of using native cli for Claude and codex is it can go off of a subscription vs an API key, but having the agent talk to an ready spun up session stored in tmux would allow for users to use the subscription model vs racking up large costs with api keys. Pair it with two subscriptions and if one 5 hour window goes blank instantly you can have a backup session research what caused the tokens to bloat without having to wait for a reset
Open Source
Integration marketplace
A place to share sub-agents and integrations other people have built, so nobody rebuilds the Stripe watcher from scratch a hundred times.
Platform
Team mode
More than one human, one shared agent. Opens real questions about whose memory is whose.
bggm21093d
Obsidian Vault integration
Integrations
To give Trillion access to your notes.
Sarah F.3d
Name and Personality
Interface
Being able to give Trillion a unique name and personality that allows the user to customize her to their communication needs or desires.
Platform
Linux support
Same story as Windows. Plenty of the people who would actually run a self-hosted voice agent are on Linux, and right now they cannot.
realtorryanlv3d
Orchistration layer
Brain
a smart routing feature that automaticly selects the best most capable model to give the best results at the lowest token cost without comprimising results. automaticly choosing and using chat cpt, codex, claude code, google gemini, etc. and all of their teirs (Terra Luna sol astra, Sonnet, opus, fable, etc.)
Thanasis L.4d
Earned Autonomy — Let Agents Earn Authority Per Action
Voice
Trillion already has human approval gates, orchestration, sub-agents, and security boundaries. I’d love to see a layer where autonomy is earned from evidence rather than enabled globally. Every consequential action would start in shadow/propose-only mode. Trillion would record what the agent recommended, what the human approved or corrected, and the eventual outcome. Authority could then mature independently per action class — observe → propose → confirm → bounded execute — instead of treating an entire agent as “trusted.” The important part is scope: trust belongs to a specific action + agent/version + tool/model basis, not to the agent globally. A material prompt, model, tool, or policy change could automatically downgrade that authority until it is recalibrated. Example: an email agent might autonomously classify inbox items after enough verified history, draft replies with review, but never send externally without confirmation. The goal is simple: don’t grant autonomy — let the system earn it.
Hamdi J.3d
Multi-languages Arabic & French first.
Agents
We can train the model to be fluent in Arabic and French
Fausto P.3d
Screen, mouse and keyboard controls.
Integrations
Can actually take over when given permission. That would allow her to complete complicated tasks like setting up programs, run web searches w/o using an API all the time. Use your frontier subscriptions to run searches etc.. Request prompts then copy and paste into coder or type questions requests into the coders chat while you're away and using sc
Zara3d
Connection to Notion
Integrations
A full Notion integration
Layer8 H.3d
Security Operator - 24/7 System Protector
Agents
The Agent check your Ports, observing your Network Traffic, outgoing and ongoing ip's. A wireshark agent which identifies live attacks.
Considering
13I've read these and I'm weighing them. One rung up from Ideas.
Platform
Windows support
Trillion runs on Mac today. Windows is the single most common reason someone cannot use her, and it is next.
lisimMost engaged4d
Maintenance mode
Platform
Trillion is a assistant who will get front loaded with lots of information, having her automatically going and doing self maintenance will help insure she stays sharp and ready to help
Open Source
One-command install
Clone, run one command, answer a few questions, have a working agent. If the first run takes an afternoon, nobody gets to the good part.
Voice
Wake word
Say her name and she wakes up, no button. The tradeoff is an always-listening mic, which is a real decision and not an obvious yes.
Voice
Interrupt her mid-sentence
Talk over her and she stops and listens, the way a person does. Harder than it sounds once audio is already streaming.
Integrations
Calendar and email
She already sees revenue and code. Calendar and inbox would let her tell you that the thing you are about to spend the day on is not the thing that matters.
Integrations
Slack and Discord
Reach Trillion from where you already are, and let her reach you there when something needs attention.
Raspberry S.4d
Automation / workflows
Voice
It would be absolutely game-changing if Tirllion could automate any given task at a scheduled time; this would ultimately result in a more Jarvis-like agent. Also, browser automation, particularly in my use case, would be life-changing.
Daniel M.Crew3d
UI colour customization
Interface
Obviously, be hugely beneficial to those with different accessibility needs, such as colour blindness, or those in differently lit environments. Plus, at the very least, it's fun customization
Voice
Custom voices
Pick how Trillion sounds, or clone a voice. Fun, genuinely useful for accessibility, and a small ethics conversation attached.
Platform
Fully offline mode
Local models, local transcription, nothing leaving your machine. Slower and less capable today, but for some people that tradeoff is the entire point.
Corey4d
PC Control
Integrations
Having the AI Assistant to be able to control your/the pc would be insanely cool, for example, “Hey Trillion, open Chrome, and go to Amazon and show me the top 10 most popular items.”, and it would do that, I personally think that that would be a great feature because it would make your workflow so much quicker.
Adam L.Most comments4d
The approval gate: anything that needs the human becomes a yes or no card on the phone, no
Interface
From six months of running my own version of this (SAMM, built for my brain, now running my whole company), the five that actually got used, in order: 1. The approval gate. Any agent that needs me files a yes-or-no card; I rule from my phone with one tap. Target is six decisions a week, not sixty. Everything below that line the agents decide and log. 2. A text line. I text her, she texts back, no app. Trillion already said it herself. 3. Desks on a schedule, one number each. Named sub-agents that fire on a clock and are measured on one number they own, not on whether they ran. 4. The room. Lights, the speaker, a spoken line when a stranger is on the site or money lands. The system has a body in the house. 5. The board. One page on the phone and the nightstand: focus next, the number, what is waiting on me, the trends. Replaced eighty slides. Happy to compare notes on camera any time. Adam, 2ndbrainway.com
Next up
1Committed, not started. No dates, because I will not hit them.
Open Source
Documentation site
Real docs. How the brain works, how to add a sub-agent, how to wire an integration, what every config value does, and what to do when it breaks.
Building now
0What I'm actually working on this week.
Already built
18Running in Trillion today. Not public yet, but it exists.
- Voice-first conversation loopVoice
- Sub-second voice responsesVoice
- Natural turn-takingVoice
- Long-term memoryBrain
- A personality that does not driftBrain
- Self-knowledgeBrain
- Sub-agent orchestrationAgents
- The agent factoryAgents