“Find my brand notes”
JEV receives candidate snippets and chooses the route. The brain should ground its response in the notes: finding a contact doesn't mean it has the answer.
Five decisions before the main brain. Use this foundation to build a Jarvis for studying, work, customer service, or personal use.

An executable core you can adapt. The standout feature is JEV Reflex via OpenRouter; you connect the brain, memory, and tools for your domain.
A structured call classifies the request before planning. It shows the model, probabilities, and reported cost.
The planner receives the tools for the selected route. If there’s uncertainty, the context is preserved.
Writes require approval by the host. A persistent record prevents repeating the same request identifier.
What’s the destination? Is the request complete? Is it for the assistant? Does it require action? Do the memory snippets answer it? Probabilities provide guidance, but don’t authorize actions.
The offline demo needs no key and installs no dependencies. To use JEV, configure your OpenRouter credentials in the environment.
The core uses native TypeScript and Node's SQLite. The real JEV call incurs API usage.
# command
node --version
git --versionStart with the demo, consult the real JEV, then connect your adapters. The example doesn't send messages to anyone.
The demo retrieves a fictional note and proposes a draft for approval.
git clone https://github.com/inematds/jarvismodelo.git cd jarvismodelo npm test npm run demo # local execution, no API
Set OPENROUTER_API_KEY in the environment without putting the secret in the code. In the INEMA installation, the example also reads the two authorized local credential files.
node examples/jev.ts --live
Check source: jev, resolved model and usage. On failure, the result reports the fallback and the process exits with code 1.
Adapt retrieve, plan, and tools as in the example. Enable the layer in the constructor:
reflex: {
enabled: true,
key: process.env.OPENROUTER_API_KEY
}Use a persistent ledger, stable IDs for each request, and argument validation. The authenticated host approves the exact proposal; the model never gets access to approve.
Want a ready-to-use interface with chat, notes, and settings? Use Jarvis v7 with JEV. Want to build another experience? Extend this foundation with the adapter contract.
Illustrative examples. The percentages for each real call may vary.
JEV receives candidate snippets and chooses the route. The brain should ground its response in the notes: finding a contact doesn't mean it has the answer.
The communication route limits the available tools. In the demo, the proposal is a fictional draft; writing only runs after approval by the host.
On an 8-second timeout or invalid contract, the assistant continues locally and reports the source. No benchmark proves “100 times faster.” The layer adds one call and needs to be measured in your use case.
Version 0.1.0. The core works; interfaces and domain connectors are up to your implementation.