echo.An open Jarvis.On every screen.
An open Jarvis you actually own. Voice, vision, memory, a translucent multi-monitor HUD. Runs on the AI subscription you already pay for. No API keys. No second bill. Cross-platform. Open source.
Made and kept.
Echo is built around three non-negotiable rules. Every design choice has to pass all three or it does not ship.
No API keys, ever.
Echo runs on whatever AI subscription you already pay for. It speaks to Claude Code, Codex, Gemini CLI, Ollama, anything that ships an official sanctioned channel. You bring your Pro plan. Echo provides nothing else to bill.
Local-first by design.
Voice loop, memory store, skills, HUD, everything sits on your machine. No SaaS in front, no proxy in the middle. Your prompts go straight from your keyboard to the brain you authorised, and back. Nothing in between.
The only cost is the one you already pay.
Echo itself is free and open source. The pricing promise is plain. If you have one Pro subscription you have a full Jarvis. If you run Ollama locally, you have a full Jarvis for free. We refuse to add a recurring fee on top.
A personal AI assistant for grown-ups.
There is no shortage of AI assistant apps. Every one of them either locks you into one provider, sends your data to a server you do not control, charges a second subscription on top of the one you already pay, or breaks the moment a model gets renamed.
Echo is the version that does not do any of that. It is a voice loop, a translucent multi-monitor HUD, a memory store, a proactive scheduler, and a bag of skills that act on your world. It dispatches each prompt to whichever AI brain you have authorised, picking the best fit by capability and remaining quota. When one is exhausted, the next takes over.
The only AI cost is the subscription you already pay. That is the whole premise.
It runs on macOS, Windows, and Linux from one Rust codebase. It uses your OS keychain for every token. It writes its memory as plain Markdown so your data is portable. It is open-source under MIT. It will never have a paid tier. It will never phone home. It is the version of Jarvis you can build today, on hardware you already own.
Authorise any AI you already pay for.
One click in the setup wizard. Echo runs the provider's own sign-in check. Your password and token never leave that provider. Echo just dispatches your prompts to whichever brain you authorised.
Claude Code
Subscription
Claude Pro or Max
Channel
claude CLI
Strong at
long reasoning, code, tool use, hooks, long context
Codex
Subscription
ChatGPT Plus, Pro, Team
Channel
codex CLI
Strong at
mixed reasoning, image generation, web search
Gemini
Subscription
Google account or Gemini Code Assist Individual
Channel
gemini CLI
Strong at
web grounding, long-context summarisation, Workspace
Ollama
Subscription
None, runs on your own hardware
Channel
localhost API
Strong at
fully offline, your own models, zero dependence on any cloud
LM Studio
Subscription
None, runs on your own hardware
Channel
localhost API
Strong at
GUI for picking models, same offline freedom as Ollama
Anything else
Subscription
Whatever the provider ships
Channel
New adapter file
Strong at
Brain trait is one file. Add a new provider the week they ship a sanctioned CLI.
Your only AI cost is the subscription you already pay.
A user with Claude Pro
Pay Anthropic for Pro. Authorise Echo to use the claude CLI. Get the full Jarvis-grade voice loop, multi-screen HUD, calendar, mail, vision, memory. No further AI bill, ever.
A user with Ollama
Install Ollama, pick a model that fits your hardware, authorise it as a brain. Echo runs end to end on your own machine. No internet, no provider, no recurring cost.
Why this matters
Most AI assistants make money by reselling API access behind a recurring fee. Echo refuses that model. The user owns the cost path. The result is the cheapest possible Jarvis on the market by structure, not by promotion.
What you actually pay
Everything Echo does for you.
The full feature list at launch. Every entry is implementable on the day Phase 8 ships. No vapourware, no maybes.
Wake-word voice loop
Say the name. Echo listens. Sub-second turn time on a midrange laptop.
Streaming TTS
Sentence-by-sentence playback the moment the brain emits a full clause. No awkward dead air.
Brain router
Capability plus quota plus user pins. Always the right brain for the right job.
Multi-monitor HUD
Glass card on the primary screen. Dashboard on the second. Ambient brief on a third. Configurable per display.
Memory across years
Plain Markdown facts plus JSONL episodes plus sqlite vector index. Survives sessions, machines, restarts.
OS keychain auth
Every OAuth token lives in Keychain, DPAPI, or Secret Service. Never on disk in plain text.
Vision when asked
Screenshot any monitor on explicit consent. Red dot indicator the whole time. Never silent capture.
Calendar awareness
Google, Microsoft, Apple. Read, create, cancel. One-click OAuth per provider.
Mail awareness
Triage, summarise, draft. Approval queue in the HUD before anything sends.
Music control
Spotify or Apple Music. Play, queue, skip, mood-based selection.
Home Assistant bridge
Lights, scenes, temperature, locks. One token, full house.
Health glance
Apple Health on macOS, Google Fit on Linux. Read-only by default.
Notes integration
Notion and Obsidian. Append, search, link. No vendor lock-in.
Code awareness
Watches your repos. Surfaces build failures, open PRs, unresolved review comments.
Proactive watches
Cron-style scheduler with quiet hours, focus mode, priority overrides.
Multi-step workflows
Borrowed from agent-orchestrator. Plan, run, replay, audit.
Single-click OAuth
Local callback server on a random port. Provider auth, then keychain. Done.
Local-first storage
Optional disk encryption keyed to OS secure storage. Forget button wipes everything.
MCP skill bus
Every skill is a small MCP server. Plugs into Claude Code, Codex, Gemini, anything MCP-aware.
Cross-platform shell
Tauri 2. Mac, Windows, Linux from one Rust core. 10MB binary, not 100MB.
Pulse animations
Listening, thinking, speaking. The HUD never leaves you guessing what state Echo is in.
Keyboard-only mode
Press a chord, type a prompt, read the reply. For meetings, libraries, noisy offices.
Subtitle mode
Live captions in the HUD. Big text, high contrast, configurable.
Web search skills
Brave, Tavily, SearXNG. Pick yours. Cached locally for repeat queries.
Image generation
Through whichever brain supports it on the user plan. Output goes to a local gallery.
Local file search
Spotlight, Windows Search, recoll on Linux. One protocol, three OSes.
Network audit log
Every outbound call by skill and host, visible in the HUD audit panel.
Skill SDK
TypeScript or Rust. Write a skill in an afternoon. Manifest, tools, OAuth scope, done.
Quiet hours respected
Outside working hours Echo is silent unless a watch is marked emergency.
MIT licensed
Same as every other open-source repo from this studio. No vendor lock-in, ever.
A different role for every screen.
Echo detects every monitor at startup and on hot-plug. Assign each one a role. The HUD scales with how you actually work.
One Rust core, three operating systems.
Local-first means actually local.
OS keychain for every token
Keychain on macOS, Credential Vault on Windows, Secret Service on Linux. Never written to disk in plain text.
Hardware-style mic indicator
Mute means mute. Bypass by skills is impossible. Red dot when capturing screen, always visible.
No telemetry of any kind
No phone-home pings, no anonymous usage analytics, no crash reports unless you explicitly opt in.
Network audit log
Every outbound call by skill and host, queryable from the HUD. You see what Echo is actually doing.
Optional disk encryption
Memory store can be encrypted at rest, keyed to OS secure storage. Toggle in settings.
A real forget button
Wipes session memory and logs with one click. A second button wipes everything Echo has ever stored.
Built slowly, shipped honestly.
Started November 2024 in private. Nine phases. Six already shipped, the last three closing now, public 0.1.0 on 1 July 2026.
Repo scaffold
Tauri 2 monorepo, CI for Mac, Windows, Linux. First glass card on the primary monitor.
MVP voice loop
Wake word, STT, brain dispatch, TTS, starter skills. Greets the user when prompted.
HUD polish
Multi-monitor detection, full glass layout, pulse animations, accessibility, subtitle mode, system tray.
Multi-brain router
Claude, Codex, Gemini, Ollama, LM Studio adapters. Scoring policy, user pins, quota awareness.
Calendar plus mail
Google, Microsoft, Apple. One-click OAuth. Morning briefings. Meeting warnings.
The senses
Vision, music, notes, tasks, health, home, Slack, news, finance. One skill at a time.
Proactive engine
Cron and watch scheduler with quiet hours, focus mode, watch templates, approval queue.
Autonomous workflows
Multi-step plans, durable orchestration, journal across restarts, multi-day tasks.
0.1.0 release
Signed installers for all three OSes. Auto-updater. Public docs. Hacker News launch day.
Boring tech, surgical complexity.
Tauri 2
Cross-platform shell, Rust core, 10MB binary
Rust
cpal for audio in, whisper-rs for STT, native APIs
React 19 + Tailwind v4
HUD windows, same stack as sarmalinux.com
Whisper.cpp
Local STT, runs offline
Piper TTS
Local TTS with OS-native fallback
Porcupine
Wake-word detection, low CPU
MCP stdio
Skill bus protocol, industry standard
sqlite-vss
Local vector store for memory recall
GitHub Actions
Three-OS build matrix, signed installers
Open by design, yours by default.
Echo exists because the pieces finally do. Voice models that run locally. Brain CLIs that respect a Pro subscription. MCP as a clean skill protocol. Tauri as a real cross-platform shell. OS keychains everywhere.
Built openly because closed assistants always end up renting your attention. Built local-first because your prompts are your business. Built MIT because the next person should be able to fork it on a Tuesday and not owe anyone a thing.
This page exists so the work is visible from day zero. The plan is published. The repo is open. When the launch lands on 1 July 2026, the only news will be that it works exactly as the page already says.
The honest answers.
Because a normal Pro subscription is meant to be the only AI cost. APIs charge per token on top, with bills that can spike. Echo refuses to put that risk on the user. If a user genuinely needs API access they can use Claude Code or Codex directly with their key set, not Echo.
Watch the build,in public.
The repo is the announcement channel. Star it, watch it, read the wiki as the phases ship. No email list, no marketing funnel, no notification spam.