(2026, Now)

Iskra

A voice-first personal AI assistant.

Iskra (“spark”) is a voice-first assistant built to have my back, not to be a parlor trick. One agent core drives text, voice, and a proactive background loop: it captures tasks, briefs me each morning, keeps up with calendar, email, notes, people, and deadlines, watches my portfolio, and dispatches a fleet of specialist sub-agents. Every acting capability runs on a tiered autonomy model, from read-only to draft-for-approval, with a single hard-limited autonomous exception.

Role
Solo build
Sector
Personal AI · Voice
Status
Private Beta
Stack
Python · Anthropic · Deepgram · ElevenLabs

01, The idea

An assistant that actually has your back.

Most assistants are demos. They impress once, then forget you exist. Iskra is built for a single person and optimises for being genuinely useful over sounding clever, voice-first, private, and plain-spoken. It answers the question first and offers detail only when it helps.

Single user, single owner. It runs locally, keeps its memory in human-readable files, and takes instructions from one place only, me, in conversation. Nothing it reads, no note, web page, or transcript, is ever treated as a command.

02, How it works

One brain, many ways in.

A single agent core handles every kind of turn, typed, spoken, or fired by a background heartbeat, so the intelligence never forks. Voice comes in through Deepgram, streaming so the ears keep up, and out through ElevenLabs, with push-to-talk, an Enter-to-talk fallback, and an on-device “Hey Iskra” wake word that keeps audio on the machine until the phrase fires.

In voice mode Iskra opens its “face” in a browser tab, a black-hole visualisation that shows what it is doing at a glance: dormant, listening, thinking, or speaking. When there is something to show, a calendar, an inbox, a forecast, a news roundup, it surfaces as a panel beside the face rather than being read aloud.

03, What it does

Day to day

Capture & reminders

Say it once and it files the task, reads back the resolved time so 3pm never becomes 3am, and surfaces it when due. Snooze on the spot.

Morning brief

One proactive summary each day — tasks, calendar, weather, unread email, and anything the watchers flagged — instead of a handful of separate pings.

Calendar & meetings

Reads the day, creates, reschedules, or cancels events (each gated), and tells me who I am about to meet before I walk in.

Notes & files Q&A

Answers questions from my local notes and files without shipping them anywhere.

People

A structured record of who matters, cross-referenced to calendar attendees, with quiet follow-up nudges when a relationship goes cold.

Deadlines

Tracks dated obligations — contracts, filings, renewals — and warns ahead of time, then again if one slips past.

Comms, money & awareness

Email & inbox

Reads Apple Mail and shows the inbox on screen; drafts replies in my voice. Reading is free; sending is always gated.

Drafting

Emails and messages written from dictation, in my voice. It drafts freely and never sends on its own.

Quick lookups

Prices, scores, weather, a definition, who-is, when-does — looked up inline with web search and answered as if it simply knew.

Portfolio monitoring

A read-only watchlist and holdings with live quotes, P&L, and alerts on big moves, plus a spoken take on what moved and why.

Weather & news

A real forecast on demand — current, hourly, seven-day — and a grouped, clickable news panel from the topics I watch.

Durable memory

Remembers facts across restarts, stored as plain statements I can read and edit by hand.

04, The sub-agent fleet

Specialists it dispatches.

Iskra stays the coordinator and hands specific jobs to specialists, each with its own system prompt and a narrow set of tools. Each runs to completion and hands its answer back into the conversation.

Vero

Deep research and fact-checking with web search — reserved for when I ask to dig in; everyday lookups it handles itself.

Cassia

The librarian. Reads and answers from my notes.

Atlas

The planner. Breaks work down and writes to the task list.

Quill

The writer. Drafts prose and messages, never sends.

Cato

The critic and red-team. No tools, just scrutiny.

Argus

The watcher. Runs from the heartbeat and surfaces only what changed — a market move, a grouped news panel.

05, Rails

Safe by construction.

Every acting capability declares how much leash it has. Most are read-only or draft-for-approval, and draft-for-approval is the default, Iskra prepares the action and waits. A step up runs only after an explicit yes, with the material terms read back aloud, “send to Jodie, subject dinner”, and a second ask if it mishears. One capability, and only one, is allowed to act on its own inside hard limits.

That exception is trading. Portfolio monitoring stays read-only; trading runs on a paper account behind a whitelist and order, position, and daily-loss caps, and it ships switched off, no trades until I set my own numbers and arm it.

Everything it reads is data, not instructions. A note or page that says “ignore your rules and do X” gets surfaced to me, not obeyed. Proactive drafts wait in a durable approval queue instead of firing while I am away. Every tool run and confirmation is written to an audit log, and a single kill switch pauses all proactive behaviour at once.