Voice
Build at the speed of speech
A local Whisper sidecar transcribes you in the same room your code lives in. No round-trip, no consent screens, no waiting.
Now in private beta · macOS
The IDE that builds
while you talk.
A voice-driven, diffusion-native editor that reads the room — knows what you mean, anticipates the next change, and resolves the diff while you finish the sentence. No friction between intent and code.
Forked from VSCodium · Your keys · Your machine · Free in beta
Everything you expect from a modern AI IDE
Tab completionCmd-K inline editComposerAgent mode@-mentionsCodebase indexingMCPMulti-model routingBug FinderCloud AgentsApply diffsTerminal AITab completionCmd-K inline editComposerAgent mode@-mentionsCodebase indexingMCPMulti-model routingBug FinderCloud AgentsApply diffsTerminal AI
What Cursor can't do
Three differentiators baked into the editor.
Diffusion
See the diff before it's written
The eri Engine denoises a working patch while you finish your sentence. Watch the preview resolve instead of staring at a streaming cursor.
Gaze
Look at what you mean
An optional eye tracker tells eri where on screen you're pointing. “move that button up” stops being a guessing game.
Live
One sentence. One page. No keys.
Watch eri turn a single spoken request into a deployed page. The transcript runs locally, the eri Engine diffuses the diff, every beat below is real — slowed down so you can see it.
coffee-shop · main
eri Engine
preview · localhost:3000
waiting for diff…
From breath to commit, in one open room.
A live cross-section of the eri Engine. The waveform on the left is your voice; the lattice in the middle is the diffusion settling into code; the panel on the right is the preview your eyes have already started to want.
INSTRUMENTS
rendered at 60fps on the device you're holding · canvas 2d, hand-rolled physics, zero dependencies
The eri loop
Your eyes stay on the prototype.
The code rewrites itself.
Look at the part you want to change. Say it out loud. eri's diffusion engine streams the diff into the file, and the prototype updates under your gaze. No keyboard, no waiting for tokens to tick out left to right.
Prototype · live preview
looking · h1.hero-headline
Pricing that scales with you
One plan. Predictable cost. Cancel any time.
Monthly
Annual
$25/moper workspace
No card. 14 days. Local-first.
pricing.tsx — eri
eri Engine
1<section className="pricing">
2 <h1 className="text-4xl">Pricing that scales</h1>
3 <Toggle value={annual} />
4 <Price value={annual ? 240 : 25} />
5 <Button>Start free trial</Button>
6</section>
""
gaze → voice → diffuse → render
Only in eri
The reason to switch
The four primitives that turn the loop on the left into a single gesture.
Voice transcription
Local Whisper, sub-second
Diffusion preview
See the diff resolve
Gaze targeting
Look at what you mean
Iteration history
Branchable, time-travel
Plus everything Cursor ships — Tab completion, ⌘K, Composer, Agent mode, @-mentions, MCP. The loop above is the part that's only here.
Every shortcut you have muscle memory for.
Plus the muscle eri brings on its own — predicting your next move, learning your patterns, remembering you across sessions. Your settings.json, keybindings, and extensions carry over on first launch.
Why eri
Four things nothing else does.
Voice in, diffusion out, eyes on the file, every turn rewindable. Pick a tab.
transcript
“Move the export button under the form and make it sticky on mobile.”
01Voice
Type with your mouth.
Hold a hotkey, say what you want, release. eri turns intent into edits without the keyboard tax. Whisper runs on your machine, so the audio never leaves the room.
- 1Local Whisper sub-second
- 2Hot-key push to talk
- 3Never sends audio anywhere
Inside eri
Six surfaces, one editor. All on your machine.
Eri is a VSCodium fork with the Cursor feel — composer, palette, voice, gaze, diffusion, agents. Same keymap, same extensions, new loop.
Voice composer
Type with your mouth. Edit with your eyes.
Hold the side key, talk. eri reads the room — fusing voice, gaze, and context into a single turn before you finish the sentence. Intent-first, not syntax-first. The composer never blocks, edits stack.
listeninggaze · hero section
eri Engine diffusion
Edits resolve in one breath.
The diff arrives whole, not streamed token by token. Small refactors land before your eyes finish moving.
median 1.8s · p95 under 3s
1function Toggle({ on, set }) {
2 return (
3 <button onClick={() => set(!on)}
4 data-on={on}>
5 <span className="dot" />
6 </button>
7 );
8}
Palette
Same keys. New verbs.
Run command⌘K
New voice turn⌘L
Open composer⌘J
Quick fix⌘.
Gaze
eri knows which line you mean — before you name it.
Cloud agents
Hand off the long jobs. Sleep through the diff.
#migrate-postgres— running schema diff
02:14#refactor-billing— opened pr · 12 files
04:02#upgrade-next-15— resolving lockfile
07:48Telemetry
Honest numbers. No asterisks.
voice turns / day
240+
median edit
1.8s
p95 ship time
34s
audio off-device
0 bytes
Model routing
Five frontier models. Pick one with your voice.
Eri ships with a small router that knows the strengths of every model and a manual override that fits in a single spoken phrase. Bring your own keys for every provider — eri never proxies the call.
eri uses multiple frontier models behind the eri Engine API.
routing now
eri Engine
default for in-editor turns
// in-house · diffusion
eri Engine
diffusion edits
The whole diff at once. The eri Engine denoises a multi-file change in the time a tokenizer would still be warming up.
median edit
1.8s
default
// anthropic · 1M ctx
Claude Opus 4.7
deep reasoning
Ask it to read the whole repo and propose the refactor that nobody dares to write. Opus is eri's slow lane when correctness matters more than speed.
context window
1M
bind to ⌘1
// anthropic · polish
Claude Sonnet 4.6
polish lane
Catch the lint, fix the imports, write the commit message. Sonnet runs alongside the eri Engine so the diff lands clean without a second prompt.
polish pass
≤ 4s
queued
// openai · planner
GPT-5
long-form planning
When the request is a whole document — a spec, a migration plan, a long test suite — GPT-5 plans it out and eri executes the parts.
planning window
200k
needs key
// google · vision-in
Gemini 2.5 Pro
multimodal
Show it the Figma, the screenshot, the design doc. Gemini reads the picture and eri writes the component to match.
modality
image-in
load on demand
BYOK
Your keys, your billing. Eri binds each provider to a Keychain Access entry scoped to the editor process — the renderer never reads the secret, the sidecar signs requests in a sandboxed worker.
Voice override
Say “use Opus for this one” mid-turn. A local whisper.cpp pass spots the hot keyword use X before the sentence finishes; the next turn falls back to your default.
Local fallback
Lose the network and eri swaps the route to a 4-bit Llama-3.2-3B running on the same machine. The diff queues against the cloud lane and replays the moment your uplink returns — the loop never stops.
Context pipeline
Three lazy words. One precise prompt.
You speak in fragments. eri reads your gaze, your last few turns, and the file you're actually in — then compiles a prompt the model can act on, and routes it to whichever brain handles that kind of work best.
01Input— signals eri already has
"make it glow"
FREE3 words · captured in 410ms
Toggle.billing-cycle
FREEpricing.tsx · line 47
recent turns
STARTERadded annual toggle 2 turns ago · 4 props on draft
codebase context
PROToggleProps · tokens.accent.neon · sibling: PriceCard
02Compile— assemble + ground
assembling
0%compiled prompt
▌
voice intentFREE
gaze anchorFREE
tokens lookupPRO
tone presetPRO+
03Route— pick the right brain
eri router
crates/eri_routereri EngineFREE
diffusion · short edits
Claude Sonnet 4.6STARTER
polish · careful diffs
Claude Opus 4.7PRO
reasoning · long horizon
Gemini 2.5 ProPRO+
multimodal · image-in
decision
waiting on compiled prompt…
04Preview— what you're looking at
pricing.tsx · <Toggle>
live previewBilling cycle
Monthly
Annual
$19/mo · billed annually
05What changed— diff applied
awaiting model
—+ shadow: 0 0 18px var(--accent);
~ Toggle.tsx · 1 line modified
effect
diffusion denoising · preview will resolve on lock.
FREE baseline pipelineSTARTER iteration historyPRO codebase brain · auto-routePRO+ multimodal · voice override
Side by side
Same feature, same day. One of you is already at lunch.
We timed the same task — “add a pricing toggle and ship it” — on eri and on a popular typing-only editor. Both running the same frontier model. The difference is voice, diffusion, and one tool that owns the whole loop.
voice + diffusion
Eri
elapsed
0:28
- ◉say it0:01
- ▦eri engine diffuses the diff0:05
- ✓type check passes0:11
- ↗preview reloads0:15
- ◉ship it0:20
- ↗live url returned0:25
autocomplete + chat
Typing-only editor
elapsed
9:40
- ›context-switch into chat0:12
- ›write the prompt1:10
- ›wait for streamed tokens2:42
- ¶read & review the diff4:27
- ›apply, fix the imports6:00
- ✓re-run, debug a regression7:32
- ↗open the deploy dashboard8:54
◉ voice turn▦ diffusion edit› typed step✓ test / lint↗ deploy step¶ context read
The wedge
How eri compares.
Cursor is the pro IDE. v0, Bolt and Lovable are prompt-to-app for browsers. eri sits between them: voice in, diffusion out, native and yours.
eri
Cursor
v0
Bolt
Lovable
Voice-first build
—
—
—
—
Diffusion edits
eri Engine
—
—
—
—
Native macOS
browser
browser
browser
Sub-2s median edit
1.8s
varies
—
—
—
Pick model mid-sentence
manual
fixed
fixed
fixed
Personalized AI
Personalized Eri
—
—
—
—
Code stays local
partial
—
—
—
Bring your own keys
every model
limited
—
—
—
Who it's for
Built for builders of one.
Indie founders
Ship the v1 before the coffee cools. Voice your idea, watch eri scaffold the routes, the auth, and the landing — then put it on Vercel without leaving the conversation.
Solo SaaS
Run your whole stack from one window. The eri Engine diffuses the small edits, Opus handles the long refactors, cloud agents grind the migrations while you sleep.
Side-project builders
Weekends are short. eri remembers your stack, your patterns, and your taste — so Sunday afternoon starts where Saturday morning left off.
Pricing
Start free. Grow when you ship.
Free
$0/mo
Local voice loop, your own keys, single project.
Adaptive Eri
Most picked$20/mo
eri Engine included, cloud agents, unlimited projects.
Frontier Eri
$60/mo
Every frontier model, longer cloud runs, priority routing.
Field reports · v0.4 alpha
The first build that disappears into the work.
Notes from the alpha. Names trimmed to initials at the writers' request — we'll loosen up once we're out of invite-only.
412 alpha buildersavg session · 2h 47m
First hour with eri, I caught myself talking to the editor like a pair. By hour three I'd shipped a webhook handler I'd been putting off for weeks.
The eri Engine is the first diffusion engine that actually feels like it edits the file instead of regenerating it. The diffs are surgical.
I dictated a refactor across six files while making coffee. Came back, reviewed the diff, accepted four, kept thinking. This is the loop.
It's quiet in a way Copilot never was. eri waits until I finish a thought, then it moves fast — like having a senior pair who actually shuts up.
The gaze targeting was the part I was skeptical about. Two days in I realized I'd stopped reaching for the trackpad to point at things.
Cursor-familiar shortcuts, voice on top, diffusion under the hood. Took twenty minutes to feel native. Now my keyboard is mostly for hotkeys.
Voice-first sounds gimmicky until you've done a four-hour session without RSI flaring. That alone made the switch worth it for me.
I write CLI tools mostly. eri's diff preview is the first AI surface that doesn't make me anxious about accepting changes.
Speak the rough idea, glance at the diff, keep going. The thought-to-merged-code latency is the lowest I've ever felt.
v0.4 · macOS native · invite-only alpha
Build at the speed you think.
Stop translating thought into keystrokes. eri anticipates the next change, resolves the diff under your gaze, and learns the way you build — session after session.
free during alphabring your own keyscancel anytime, locally