Five decisions before the response: JEV identifies the destination, evaluates the request, and checks evidence from the notes before the main brain responds. Probabilities, latency, and fallback are visible.
Real JEV, via OpenRouter, evaluates five questions in one call: destination, complete request, direction to the assistant, need for action, and coverage of the notes.
In Settings → Connections, configure your credential. Using JEV consumes API credits and sends the request and up to six selected excerpts.
In Preferences, enable “JEV · Reflex routing” and save. Then ask a question in the chat.
Expand Reflex in the chat: model, probabilities, latency, and reported usage. If it fails, the chat continues and indicates the local fallback.
Probabilities are not authorization to act. The chat still doesn't send emails or manage your calendar. We don't promise “100 times faster”: that needs to be measured in your use.
Create another Jarvis with this foundation + JEVThis page is the public guide. The application runs on your computer, with your own accounts.
Import MD, TXT, and text-based PDFs. Review notes, save memories, and consult the excerpts used in responses.
Codex and Claude use OAuth from the official runtimes. OpenRouter uses an API. Switch brains for each conversation.
Enable Kie media, browser voice, screen frame, and manual focus. Local search works without generative AI.
You control what you save and which provider receives the selected context. The database stays in your local installation.
Node.js 24, npm, and Git. Tested on Linux. Codex, Claude, and pdftotext are optional, depending on the features you choose.
Use Node 24. After installing the dependencies, diagnostics check the runtimes without printing credentials.
# command
node --version
npm --version
git --versionThe commands below run in your terminal. The GitHub Pages site does not host the server or receive your credentials.
Clone the public code and build the application.
git clone https://github.com/inematds/jarvisv7.git cd jarvisv7 npm ci npm run preflight npm start
Go to http://127.0.0.1:4700. Choose a profile; the fictional notes are optional. Don't expose the port to the network.
In Settings → Connections, use the Codex login in your browser or your existing login. For Claude, authenticate the runtime in the terminal and enable the personal adapter.
claude auth login
In Brain, select a provider, model, and effort. Start with medium. OpenRouter and Kie use their own credentials, saved only in this installation.
Import documents into Knowledge. Type “Remember that…” to save a memory. The provider button in the message field switches the conversation’s brain and sends the limited history to the next provider.
In Preferences, enable media, voice, screen, or focus. With Kie, confirm the usage before generating. Screen sends only one explicit frame with the question; focus is manual.
Stop the server with Ctrl+C and copy the database and library. The interface’s JSON export does not include credentials; the web import only adds notes and memories.
npm run backup -- --stopped
Configure the releases repository in the environment before starting. To install a future tag, stop the server, keep Git clean, and use the command below. v0.1.1 is just an example.
export JARVIS_RELEASE_REPO=inematds/jarvisv7 # Instalar uma tag quando ela existir: npm run update -- v0.1.1 --stopped
Use one of your notes or the optional fictional examples from the initial assistant.
Remember that project Aurora will launch in October.
Jarvis saves a real memory, available for review under Memories.
What do my notes say about project Aurora?
Search retrieves excerpts. With a connected LLM, the response can cite sources; in local mode, it shows only the excerpts.
Jarvis v7 is the project name. 0.2.0 adds JEV Reflex to the initial foundation, with documented limits.