Built for privacy-first support teams · macOS · QVAC SDK

The local AI assistant for privacy-first support teams.

A 5 MB macOS app that drafts replies, polishes tone, translates across 7 languages, and queries your internal docs — entirely on your laptop. No cloud APIs. No accounts. No telemetry. Powered by QVAC.

macOS · Apple Silicon·QVAC local inference·No accounts, no telemetry

Hard facts, not marketing stats

100%
On-device · QVAC
0 / 0
Accounts · Telemetry
7
Languages supported
~5 MB
Signed .dmg

The problem

Support teams shouldn't have to choose between speed and privacy.

Three pains every support lead handling sensitive data already knows.

Every minute of "let me check" costs you the ticket

Agents tab between Zendesk, Slack and three Confluence pages to draft one reply. CSAT drops, queues pile up, and the customer churns before the answer arrives.

Cloud AI doesn't know your domain's vocabulary

Generic copilots hallucinate policy, miss your tone, and ask the customer for "your deposit ID" instead of the TXID, address and network — or for "your case number" instead of the policy ID, claim type and date of loss. Your senior agents end up rewriting every draft — so the "AI assist" actually slows them down.

And the worst one

Every ticket you paste into ChatGPT is a data leak waiting to happen

Customer emails carry PII, payment details, health info, internal credentials. Cloud LLMs log it, train on it, and put you one breach away from a GDPR/HIPAA incident. So Legal blocks the tool — and your team is back to copy-pasting from docs while competitors ship faster.

The five things Cortex does

Five modules that ship today —all running on the agent's laptop.

Chat Agent

Paste a customer message, get a ready-to-send reply that already speaks your domain — whether it's a TXID and memo/tag, a policy ID and claim type, or an account suffix and dispute reason. Streaming output, session history, ⌘N for new chat, Esc to stop.

Grammar & Style Check

Polish any draft to your team's voice — facts, numbers and names stay intact. Temperature capped at 0.3 so it polishes, never rewrites.

Smart Translate

Translate between EN · ES · FR · PT · DE · IT · ZH preserving technical terms and support tone. One click to swap source ↔ target.

Response Templates

Six starter ticket types that ask the right follow-up questions out of the box — fully editable to fit your domain. Add ticket-specific context and the draft incorporates it.

  • Withdrawal Issue
  • Deposit Not Credited
  • KYC / Verification
  • API / Integration
  • Security / Account
  • General Acknowledgement

Local Knowledge Base (RAG)

Point Cortex at a folder of PDFs, Markdown or plain text. QVAC builds vector embeddings locally. Citations on/off. No web crawling, ever.

Domain-fluent

Speaks the language of your queue.

Most LLMs treat your industry's jargon as generic text. Cortex respects the terms that actually matter — TXIDs, memo/tags, policy IDs, claim types, account suffixes — and asks for the right artifacts on the first reply.

Crypto

TXIDmemo / tagnoncedeposit address2FA resetERC-20 · BEP-20 · TRC-20gas feewithdrawal whitelist

Fintech & Finance

policy IDclaim typeaccount suffixIBAN / SWIFTdispute reasonchargeback codeKYC / AMLsettlement window

Same ticket

Hi, my withdrawal hasn't arrived yet. Can you check?

Generic cloud AI

Hi! Sorry to hear that. Could you please share your withdrawal ID so we can look into it? We'll get back to you as soon as possible. 😊

  • · Vague "withdrawal ID"
  • · Misses network, memo, address
  • · Emoji breaks tone rules
Cortex

Thanks for reaching out. To trace this withdrawal I need: the TXID from your transaction history, the destination address, the asset and the network (e.g. ERC-20 vs BEP-20), and the time you submitted it. If the network requires a memo or tag, please include it as well.

  • · Asks for TXID, address, asset, network
  • · Knows memo/tag is per-network
  • · Direct, no fluff, no emoji

See it run

No fake demo. Real app, real .dmg.

A 45-second walkthrough: paste a ticket, get a streaming draft, pick a template, switch the model — all offline, on a stock MacBook.

Screencast lands here — coming with v0.1 launch.

cortex-walkthrough.mp4

01
Paste a ticket
Customer message → streaming draft in seconds.
02
Pick a template
Withdrawal, deposit, KYC, API — pre-shaped.
03
Send it
Copy to clipboard, paste into Zendesk, done.

Privacy by architecture

Hardened by architecture, not policy.

Most "AI support tools" promise privacy in a Terms of Service. Cortex can't leak your data because the wire to leak it over doesn't exist. Read the list — every item is a structural fact, not a setting you can toggle off.

No cloud API calls

Model weights live in ~/.qvac/models. Inference runs on your CPU/GPU.

No accounts, no auth, no login screen

There is nothing to sign up for. Open the app, start drafting.

No telemetry, no analytics

No crash reports phoned home. No usage pings. Nothing.

RAG reads only the folder you choose

No web crawling, no auto-sync, no background indexing of your disk.

Works offline

On a plane, on hotel Wi-Fi, behind a corporate firewall, in an air-gapped lab.

Open file format for settings

Export the full configuration as plain JSON. Audit it. Diff it. Sign it.

For team leads

Standardize 50 agents in one file.

Cortex is a single-user app by design — but it ships with the primitives you need to roll it out across a whole shift without a config server, an admin panel, or an IT ticket.

Export / import settings (.json)

A lead configures the system prompt, tone rules and KB folder once, exports the file, shares it with the team. New agents import in one click and sound identical from day one.

agent-prompttone-ruleskb-pathmodel-id

Tone presets + fine-grained toggles

Four presets — Professional, Concise, Detailed, Empathetic — plus individual switches: no emojis, direct-but-polite, prioritize security warnings, max reply length 60–1200 words, plus free-form extra instructions appended to every prompt.

no-emojisdirect-but-politesecurity-firstmax-words

Bring your own model

Three pre-configured lightweight models with one-click download, or paste any local GGUF path on disk. Temperature and max-tokens sliders for power users.

Llama 3.2 1BQwen3 1.7BQwen3 4Bany local GGUF
Configuration travels with the file. No server, no sync, no account.

Under the hood

Built on the fastest local stack.

Core
QVAC SDK
Tether's local inference runtime
Backend
Rust
Memory-safe systems layer
Shell
Tauri 2.0
Native desktop shell
Frontend
React 19
UI with concurrent rendering
Distribution
~5 MB .dmg
Signed installer · Apple Silicon
Weights
GGUF models
Llama 3.2 · Qwen3 · BYO
macOS · Apple Silicon recommended

Download Cortex.

Signed .dmg, ~5 MB. First chat triggers a one-time model download (~0.5–3 GB depending on the model you pick). Cached in ~/.qvac/models — instant on every launch after that.

Download Cortex v0.1.0 (.dmg)

No account. No telemetry. No subscription.

Get notified when v0.2 ships