echo v1.35.1 · one customer per machine
Echo logs into your servers and picks up where you left off.
An AI operator on a VM you own, running on your keys. It reads the repo and your last session, restarts the work that stalled, and messages you when something needs your call.
$ curl -fsSL \ https://echo.yourco.com/hooks/mesh/enroll.sh \ | sudo bash → authorizing key for echo → sshd port 1979 detected → firewall opened to 1 address → ssh handshake proved ✓ registered · reading the repo now
billing-api
- 18 files uncommitted on main
- retry queue half-built, no dead-letter yet
- last commit 11 days ago
- 3 failing tests, all in signature checks
One line on a server, and it knows the project.
Paste the enroll command on any box you own. The script authorizes a key, finds your sshd port, and opens the firewall to your Echo's address and nothing else. A couple of minutes later there's a card on your dashboard describing what you were last building on that machine, what's uncommitted, and what broke.
It wrote that after reading the git history and your own past Claude Code sessions on the box.
You didn't tell it any of that.
Your VM, your keys, your bill.
Every Echo is one customer on one machine with its own domain. You paste your API keys into the dashboard and usage bills to your provider account at their price. We don't hold a key for you and there's no markup on top.
It still updates like a product, not a script you installed once. Thirty-five releases so far, nineteen numbered migrations, and three channels: stable, beta, edge. Releases go out in waves, and wave zero is our own machines. A bad one reaches us before it reaches you.
Your hardest questions get argued over before they get answered.
Switch the model picker to The Council and the question goes to three frontier models at once. Sonnet 4.5, DeepSeek v4 Pro and GLM-5.2 each reason on it separately. Opus 4.8 reads all three and writes the answer.
You see every take before the synthesis, so when two of them disagree you find out instead of getting an average.
Cut over in one window. The dual-write path is the risk, not the schema.
Disagrees. Backfill 90 days first or the reconciliation job never converges.
Cut over, but only after the backfill. Both of the above, in order.
Backfill first. Two of three flag reconciliation as the failure mode, and it's the one you can't roll back.
Worth it for a migration plan, a contract, or a bug you've stared at for two days. Overkill for checking your calendar.
It costs roughly four times a normal answer, on your key, so it sits behind a switch instead of running by default.
Echo answers the phone and sits in the meeting.
Point one of your numbers at Echo and it picks up in well under a second. You draw the branching in an editor, and a menu reads keypresses and spoken intent in the same node. It checks your calendar for real openings before offering one, transfers when it should, and posts a signed summary to your webhook the moment the call ends. Flows are validated when you save, so a broken one never reaches a caller.
In meetings it joins the room as a participant. Captions run live, the transcript and notes land in your inbox when everyone leaves, and it answers when you say its name instead of talking over you. Start a recording and every person in the room sees the indicator.
There's a photoreal face if you want Echo on camera. The reasoning stays on your machine and the vendor only renders the head.
What you actually get.
| Version | v1.35.1 · 35 releases · 19 numbered migrations |
|---|---|
| Machine | 4 cores, 4 GB, no GPU. That's what ours runs on. |
| Chat | A persistent agent process, so there's no cold start on every message |
| Reach | Dashboard, Telegram, WhatsApp, iMessage, email, SMS, phone |
| Memory | An Obsidian vault you can open, edit and clone with git |
| Media | Stills and motion generated on your own kie.ai key |
| Mesh | As many servers as you own, one command each |
| Models | Anthropic, OpenAI, Moonshot, DeepSeek, Vertex or OpenRouter. Switch without redeploying. |
Two things to know. Setup happens with you on a call, because there's no self-serve signup yet and we'd rather say so than put a dead button on this page.
And Echo runs on your API keys, so you'll need accounts with whichever providers you want it using. That's what keeps the usage on your bill and your data out of a shared pool.
Book fifteen minutes.
We provision your VM while you're on the call. Bring the API keys you want it running on. You leave with a dashboard on your own domain and your first server enrolled.
Your operator, on your machine. Dashboard, voice, chat, automations, phone, meetings, and as many enrolled servers as you own.
Everything in Standard, plus The Council on the questions you don't want one model answering alone.