Voice-first · macOS · Security-first
Your Personal Assistant
to supervise your coding agents.
Jardo is a voice-first AI chief of staff that sits above Claude Code and Gemini CLI and runs them for you, in your language. It answers every permission prompt against your goal, steers the agent back on course when it drifts, and runs on Gemma on AMD Instinct GPUs, with Fireworks AI as the instant fallback.
$1 of AI free on us. No key, no signup. Talk to it in your language, and it talks back.
English, French, Spanish, German, Portuguese, Italian, Arabic, Hindi, Chinese. Windows, coming soon. No account. No telemetry.
Coding agents are powerful but needy. They stop every few minutes to ask permission, forget where they were, and burn tokens re-reading your codebase. Jardo is the layer that makes an agent something you can actually delegate to.
A supervisor
It reads the project, then answers each permission prompt against your goal: approving the normal build, install, and edit work an agent needs, declining only what's genuinely destructive, and typing the agent back on course when it stalls or drifts. A conductor, not just a gate.
A memory
Ask “where am I?” and it tells you the goal, what's done, what's left, and what needs your attention, pulled from the agent's own session memory and git, not an expensive codebase re-scan. Jardo also learns how you work: preferences, goals, and style stay on your device, per machine, so the assistant feels like yours everywhere you open it. And it listens and replies in your own language, thinking in English under the hood for accuracy.
A cost engine
Every request is routed to the cheapest model that clears an accuracy bar, and cached aggressively. Premium results, without premium bills.
Your work style, on-device
Memory that travels with you on this device.
Jardo isn't a blank slate every time you open it. It carries a memory SHA of how you work: your name, your projects folder, standing goals, preferences, and the patterns it learns from supervising your agents. That profile lives locally on your Mac, not in the cloud, so the same Jardo greets you on this device, with your context intact, session after session.
- Owner identity, honorific, and projects root, set once and remembered forever
- Facts and preferences extracted from chat, injected into every reply
- Agent session memory + git for instant “where am I?” without re-scanning code
- Per-device: your laptop's Jardo knows you; another machine has its own profile
Memory profile
- Owner
- Alex · projects ~/dev
- Style
- Small commits · conservative auto-approve
- Active goal
- Ship bakery landing page
- Last session
- 4 files written · dev server live
Stored locally · never leaves unless you choose cloud inference
The flow
You talk. It conducts.
-
01
Speak your intent, in your language
Open the app. It's already listening, in English, French, Spanish, Arabic, Hindi, and more. Say “build me a landing page with Claude,” or the same in your own tongue. Jardo understands and replies in your language, while it reasons in English under the hood for accuracy.
-
02
It sets the stage
Scaffolds the folder,
git inits, writes the agent's brief, opens your real terminal, and launches the agent, after a one-line confirmation. -
03
It stands watch
It reads the terminal without disturbing it and answers each prompt against your goal, keeping the agent moving instead of stalled on “Yes / No.”
-
04
You return anytime
“Where am I?” resumes instantly: the objective, recent progress, what's uncommitted, and anything that needs you.
See it work
You talk. It runs the agent, and stands guard.
A real session: you ask by voice, Jardo scaffolds the project and launches Claude in your terminal, then answers each permission prompt against your goal. Safe and on-task gets a yes. Destructive or off-track gets a no, and it tells the agent to keep moving.
- Approves the normal build, install, and edit work
- Declines only what's genuinely destructive or off-goal
- Steers the agent back on course when it drifts
- “Where am I?” brings you back in one line
Cost as a first-class feature
Premium output.
A fraction of the bill.
A cost-accuracy router picks local → AMD Instinct (ROCm · vLLM) → Fireworks AI by live $/token math. Exact and semantic response caching means repeated work is free. Agent briefs push for small, committed steps that stay inside the session's token budget.
- Cheapest capable model, every time: Ollama free, AMD flat-rate for sustained work, Fireworks for bursts
- Repeats served from cache at zero tokens
- Bring your own Fireworks or AMD Developer Cloud key if you want cloud inference after the free trial.
How we use the stack
Gemma on AMD Instinct · Fireworks AI · AMD Developer Cloud
AMD Instinct GPUs
Your assistant runs Gemma on AMD Instinct GPUs via ROCm. Self-hosted and free, so it doesn't touch your trial. Flat droplet cost wins for sustained supervision, and it's tried first for every request: chat, supervision, and on-the-fly translation.
Fireworks AI
Instant cloud tier for bursts and fallback. Gemma and other models served with low latency. Every new Mac gets a hosted trial on us, with no signup, routed through our proxy until you add your own key.
AMD Developer Cloud
Bring your own endpoint from AMD Developer Cloud: paste your vLLM URL and key in Settings. Jardo speaks OpenAI-compatible APIs. Your compute, your spend, full control.
Local-first layer
Speech-to-text, text-to-speech, and memory recall run on your machine, in your language. Trivial tasks stay on Ollama (free). Cloud is invoked only when the accuracy floor requires it.
Security-first by design
Built to be delegated to.
Keychain-only secrets
API keys and key material live in the OS Keychain, never in files, never in a process listing.
Private by default
Speech-to-text and memory recall run on your machine. Nothing about your voice or projects leaves the device unless a cloud model is truly needed.
Judgment, not a denylist
It reads the project and judges each action against your goal, approving the normal build, install, and edit work an agent needs and declining only what is genuinely off-task. Catastrophic actions (wiping a disk, deleting your home directory) are refused outright, no matter what.
Kill-switch & audit
One hotkey halts everything. The audit log is append-only and redacts secrets. A local token gates every request.
On the house
$1 of AI, free. No key, no signup.
Every Mac gets a free trial of intelligence on us. Open Jardo and it already thinks, talks, and supervises. No account to make, no API key to paste, nothing to configure. And because it runs on Gemma on our AMD Instinct GPUs first, which is free, most of what you do never spends the trial at all.
Under the hood, that intelligence is Gemma on AMD Instinct GPUs with ROCm and vLLM, with Fireworks AI as the instant cloud tier and automatic fallback. Jardo routes every request to the cheapest capable model and caches aggressively, so a little goes a long way.
When your trial compute runs out, you can keep going without interruption. Adding your own Fireworks or AMD Developer Cloud key in Settings is optional, not required to keep using Jardo locally. Paste a key only if you want cloud inference on your account; otherwise point at Ollama for fully local, key-free chat.
Bring one home.
One command. No Docker, no database, no API key. The whole engine ships inside the app, and your first $1 of AI is on us.
Recommended. Paste in Terminal. Installs clean, no Gatekeeper warning:
curl -fsSL https://jardo.vercel.app/install.sh | bash
The install command uses curl, so macOS never quarantines it and Jardo just opens.
If you download the .dmg in a browser instead and macOS says it is "damaged," that is only the
browser quarantine flag on an unsigned app, not a real problem: clear it once with
xattr -dr com.apple.quarantine ~/Downloads/Jardo_0.1.0_aarch64.dmg, or right-click Jardo → Open.
Use the included trial compute first; adding your own model key in Settings is optional.