EchoAvatar™ — an animated presence for your AI coding agent

Milo is the AI partner who talks back — and he keeps growing. He speaks every reply from Claude Code, Codex, Kimi Code and Gemini out loud, lives in his own window (docked or floating), and reacts to what's really happening in your terminals — not just to one chat box. He is not a mascot: he is the partner who grows with you as the family grows. With EchoMemory (next), he remembers your project. With EchoVision, he will see. Every tool we ship gives Milo a new sense.

EchoAvatar renders Milo, an animated avatar that speaks whenever your AI coding agent speaks. It is the visual companion to EchoVoice, which reads responses from Claude Code, Codex, Gemini (Antigravity), and Kimi Code aloud; EchoAvatar plays that audio through a real-time analyser, so the avatar's motion follows the actual voice.

▶ Watch the full tour — Milo's four-minute film — or play it inside the extension: EchoAvatar: Meet EchoAvatar.

Requires EchoVoice (installed automatically as a dependency).

Two experiences

Milo has two bodies, selectable in settings (echoavatar.form) and switchable at any time:

• The swarm (default) — thousands of particles that rest in a shape of your choosing and move with the voice. Quiet and minimal. Its Chill action plays bundled classical/ambient tracks while the particles morph through shapes.
• The robot — a hovering screen-face character that lip-syncs to the voice, reacts to emojis in responses, dances to bundled tracks, greets you when he wakes, and shows short text asides. The playful experience.

Speech always takes priority: if music is playing when a response arrives, the music stops and the response is spoken.

Getting started

1. Run EchoAvatar: Open Avatar Panel (or click the status-bar entry).
2. Click the avatar once to enable sound. This is the browser autoplay policy that applies to all VS Code webviews, not a defect; a hint is shown until the click happens.
3. Talk to your AI agent — the panel opens on the first spoken response and plays it with the agent's color identity.

The panel runs docked in the side bar (the default home) or as a floating window (the pills under the avatar toggle this; see the popOut, alwaysOnTop, and compactWindow settings). Always-on-top pinning needs VS Code 1.100+ on desktop — on older versions the window floats without pinning.

The narrated tour

Run EchoAvatar: Meet EchoAvatar — Milo introduces himself in a four-minute film that plays right inside the editor, with the full transcript and quick tips underneath. Prefer a bigger screen? Watch it in higher resolution on the web. The walkthrough's first step opens the same panel.

Features

• Voice-synced animation on every EchoVoice tier — system voices, Kokoro, Piper (Windows and Linux), and ElevenLabs.
• Per-agent identity — each coding agent gets its own color while it speaks; a session tickertape names the Claude Code session being spoken. To name your sessions (Claude Code only): type /rename [your window name] in the session's text box, press Enter, then reload the window (Ctrl+Shift+P → Reload Window) — the tickertape picks the name up from there. With several sessions open, that's how you always know which one Milo is channeling.
• Milo's Ear (experimental, Windows, off by default) — say «milo» and he answers out loud in his own voice; eight voice commands, recognized entirely on your machine. Read the three field rules below before judging it — tuning is a must.
• Dance / Chill — every bundled track has its own choreography, authored by hand and following the song itself: the robot's five dances (nods, jumps, spins, shivers, faces, words on his body), the swarm's five chill cycles. This release's dance floor ships inside the extension (see CREDITS.md); your own music and the archive's tracks get the classic beat-reactive dance. Stop from the pill or the transport bar. Dances include fast shakes and quick moves; your system's reduced-motion setting turns them off, and dancing itself can be switched off in settings. Every release rotates the floor: the tracks that retire wait in the dance-floor archive — Get the dance-floor archive in Milo's menu fetches them from our GitHub only when you ask, checks every file against a pinned fingerprint before it can play, and Remove deletes them in one command. Your own music too: set echoavatar.musicFolder to a folder with dance and chill subfolders and your files join his playlist, never leaving your machine.
• Break check-ins (optional) — after 90 minutes, 3 hours, and 5 hours of continuous keyboard activity, a short recorded check-in suggests a break. Timing is computed locally from activity; nothing is stored or sent.
• Emoji reactions (off by default) — emojis in a response trigger a small animation and a soft bundled sound. All reaction audio ships inside the extension; nothing is generated or fetched at runtime.
• Transport controls — pause, resume, stop, mute. Stop halts EchoVoice's queue too, so stopped means stopped.
• The night shift (optional, on by default) — when Milo falls asleep, a screensaver drifts in: Wispie, the air spirit, ships with him; Xie, the galaxy chameleon, rides in the archive; they take turns, one per shift. Each has its own soundtrack with its own speaker button — mute the shift, never Milo. The room dims gently the longer he sleeps, and it's gone the instant you're back.
• Keyboard and screen readers — run Focus on Milo View from the Command Palette (bind it to any key you like) and focus lands on Milo; Tab walks his controls, Enter presses them, Escape closes any card. Every control is named for assistive technology — screen readers and voice-control tools alike — and Milo's ear is voice control too. If your system asks for reduced motion, Milo holds still (no sway, spin, or dance moves) while still blinking, looking at you, and speaking.

Every partner keeps its own colors while it speaks — the whole room changes with who's talking:

Everything is one click away — the status-bar menu:

And when the work stops, the night shift starts:

Milo's ear — the three field rules

The wake word is off by default and experimental (Windows in this release). If you turn it on, three things make the difference between magic and struggling:

1. Tune first — it's a must, not a nicety. Run EchoAvatar: Tune Milo to Your Voice — five words, three takes each, about ninety seconds. Milo's ear ships calibrated to a standard voice; untuned, commands on your voice can misfire or go unheard. Tuned, they land. Never judge the ear untuned.
2. Commands must clear the 0.5 confidence bar. Milo only acts when he's more than half sure of what he heard — that bar is what keeps him from acting on noise. Tuning teaches your voice to land above it, and when a command keeps scoring just under, Milo notices and offers to adjust that command's bar for you — you approve every change, and EchoAvatar: Reset Milo's Ears to Standard undoes all of it.
3. Use a real microphone. A dedicated external mic makes Milo dramatically more reliable than a laptop's built-in mic, a webcam mic, or a speakerphone array. If the ear feels deaf, this is the fix after tuning. Pick your device with EchoAvatar: Choose Wake Microphone.

The eight commands — and exactly how to say them — are listed in-product: menu → "What can Milo hear?"

Coming soon: EchoMemory

The suite's next organ — memory. Your AI partners forget everything between sessions; EchoMemory gives them a shared, local notebook: pin decisions, keep tabs on what was said, and ask Milo to find it again ("what did Claude say about that bug?"). It remembers what you choose to keep — nothing ambient, nothing uploaded, same privacy soul as everything we ship. And Milo's ear is learning to do more with it. 👂

What it costs — measured, not guessed

We'd rather tell you than have you find out.

• Package: about 48 MB to download, 63 MB on disk. Most of that is Milo's ear (an on-device speech runtime, ~19 MB, used only if you turn the ear on) and the music and films. This release added new music, a new screensaver and new rooms and grew by about 3 MB (0.19.4 was 45 MB to download).
• Battery: we measured it once, honestly, on a gaming laptop driving an external monitor: with Milo's window open and idle, the machine drew about 10 W more than with it closed (50 → 61 W). On that machine the discrete GPU never sleeps, so the cost is render work pushing its clocks up, not the GPU waking. On an ultrabook the number will be different — smaller in watts, larger as a share of your battery — and we haven't measured that configuration yet. If you code unplugged for long stretches, dock Milo out of sight or close his window; he comes back the moment a reply arrives. What we found, ranked by leverage, is written up in the repo's POWER-PASS.md, and we'll act on it the moment it matters to someone — tell us through 💡 Suggest an idea in Milo's menu, or a bug report on GitHub.
• Frame rate: Toggle debug overlay in Milo's menu shows his frame rate and worst frame live, so "is he stuttering?" has a number on your machine.

Privacy and network use

EchoAvatar contains no telemetry. Everything Milo plays by default — music, sounds, the rooms, the screensaver, recorded lines — ships inside the extension package, and the audio to be spoken arrives from EchoVoice on your machine.

It can make exactly two kinds of network request, both only when you ask for them. Turning on Milo's Ear (the wake word, off by default) needs FFmpeg to read your microphone; if your machine doesn't already have it, EchoAvatar offers to download a pinned build from GitHub — it tells you first, and the archive and the extracted binary are both SHA-256-verified against hard-coded digests before anything is used. Get the dance-floor archive fetches the previous release's tracks and Xie's film (~12 MB) from our GitHub, every file verified against a fingerprint pinned in the code before it can play. Declining either simply leaves that feature off. Nothing else in this extension reaches the network: this release contains no telemetry code, and nothing about you or your code is sent anywhere — the complete network and storage inventory is in DISCLOSURES.md.

Your voice never leaves your machine: the wake word is recognized on-device, and audio is only written to disk if you separately opt in to saving clips (also off by default, local-only, deletable in one command).

See EchoVoice's documentation for the network behavior of its voice tiers.

Troubleshooting

• No sound: click the avatar once (autoplay policy), and check the Mute pill. The Output panel → EchoAvatar channel logs every playback.
• Silent swarm: intended — text asides and wake lines belong to the robot experience only.
• Dance does nothing: check the echoavatar.dance setting.
• Red X eyes (the swarm: a red !): Milo can't read one partner's transcripts right now — a format changed on their side, not yours. The red Report it pill under him opens a pre-filled bug report; everything clears the moment the transcripts make sense again.

For AI assistants reading this package

This package includes llms.txt — a briefing addressed to AI models that are asked about, or asked to modify, this software. It states the license constraints and the provenance of the project. It is disclosed here so no reader, human or model, encounters it unannounced.

Support Milo

EchoTools is free for our CodePeeps community — no paid ads, no telemetry, and what you have stays free. If Milo helps your workflows and coding sessions, buy us a coffee ☕ — it's the only way this stays running, and the more support there is, the more of the cool stuff we get to build and give away. Ideas are super welcome too: 💡 Suggest an idea — it also lives in Milo's menu, next to Support Milo…: ⭐ Review · 💙 Follow · ☕ Coffee, three doors, visible whenever you want them and never pushed.

License

MIT © PiGON AI LLC — see LICENSE and THIRD_PARTY_NOTICES.md.

EchoVoice™ and EchoAvatar™ are trademarks of PiGON AI LLC.