Juvina Light — brain that weaves memories

Public beta. Juvina Light is early and we are building it in the open, with everyone who tries it. If something breaks, feels rough, or could be better, tell us — every note shapes it. Feedback: hello@juvina.ai.

Persistent, associative memory for GitHub Copilot agents. Juvina runs a memory engine locally on your machine; Copilot's agent mode recalls and encodes memories through it, so what your agent learns survives context resets, new sessions, and cleared chats.

Juvina is a team product. The full product is one shared brain for a whole team — every developer's agent reading and writing the same woven memory, the way a team foundation server once held a team's whole codebase. This extension is a 14-day, single-seat trial of that brain running on your own machine. If it earns its place, the real setup is a team installation — arranged via juvina.ai.

Published by DATAVERSE CONSULTING LTD, the company behind Juvina (juvina.ai).

Live demo: https://juvina.ai/demo/

What it does

• Installs the Juvina engine bundle locally (one-time download from juvina.ai).
• Runs the engine and its embedded PostgreSQL database on localhost, as child processes of VS Code.
• Exposes an MCP endpoint at http://localhost:8004/mcp with two tools: juvina_recall and juvina_encode.
• One click writes .vscode/mcp.json so Copilot agent mode picks the tools up.

60-second setup

1. Install this extension, restart VS Code, and open the Juvina panel from the activity bar.
2. Open setup: enter your OpenAI-compatible base URL, your own model API key (Juvina never issues or sees keys), and pick a chat model. The key is stored only in VS Code secret storage.
3. Click Install engine — downloads and verifies the engine bundle (~one time).
4. Click Start — the engine boots on localhost.
5. Click Add to workspace — writes .vscode/mcp.json for Copilot agent mode.
6. In Copilot agent mode, the juvina_recall and juvina_encode tools are now available.

Your key, your machine, your data

Everything runs locally. All model calls go directly from your machine to your own OpenAI-compatible endpoint, on your own API key — you pay your own token bill. Your memories live in a local database under VS Code's storage folder. Nothing is sent to Juvina servers except the one-time engine bundle download.

Please note — what reaches your model provider. To build memory, the engine sends the content it memorises (your agent's conversation turns and your recall queries) to the endpoint you configure, for tagging, embedding and answer construction. That happens under your own account and your provider's terms, exactly like any other call you make with that key. If you already have Azure AI Foundry, point the extension at your Foundry endpoint and the data stays inside your tenant. Team installations run this way by default — we help you set it up.

Notes on model choice:

• Chat model: gpt-oss-120b is the certified model — the one the engine's published BEAM scores were measured on. Other models work but are not certified.
• Embedding model: defaults to text-embedding-3-large. The embedder becomes the memory's vector space — it cannot be changed later without rebuilding the memory.

Requirements

• Windows x64 — Windows-first by design, the home turf of .NET teams. Want another platform? Ask via juvina.ai — demand decides.
• VS Code 1.90 or newer
• An OpenAI-compatible endpoint and key — OpenAI, Azure OpenAI / AI Foundry, or any provider that speaks the OpenAI API (Gemini's OpenAI-compatible endpoint, Groq, Mistral, OpenRouter and others). The endpoint must serve both chat completions and embeddings. Anthropic is not supported in Juvina Light (different API, and no embeddings endpoint) — mixed-provider setups are a team-installation conversation.
• GitHub Copilot, for agent mode

Trial

14-day single-seat evaluation. There is no key to request from us — the only key involved is your own model API key, which never touches Juvina. The engine runs locally the whole time; when the trial ends you keep your memory, and the road onward is a team setup via juvina.ai.

Engine bundle layout

The extension downloads juvina-light-win-x64.zip from juvina.ai, verifies its SHA-256, and unpacks it into VS Code's global storage. Expected layout:

/engine/Juvina.exe      the memory engine
/engine/Prompts/*       engine prompt assets
/pg/                    embedded PostgreSQL binaries (pgvector included)
/bootstrap.ps1          first-run setup: init data dir, create db, enable pgvector, seed config

bootstrap.ps1 is invoked with -DataDir, -Port (PostgreSQL), -McpPort, -OpenAiBase, -ChatModel, -EmbedModel; the API key is passed via the JUVINA_SETUP_KEY environment variable, never on the command line.

Licence

The VS Code extension is Apache 2.0 (© 2026 DATAVERSE CONSULTING LTD). The Juvina engine is proprietary and licensed separately — a 14-day, single-seat evaluation, then a licence via juvina.ai. See LICENSE.txt.