Pre-alpha · open source

The AI assistant for Linux that never hides what leaves your machine.

Veya is an open-source, privacy-transparent, system-wide AI assistant for Ubuntu/Linux — an Apple Intelligence alternative. One intelligence daemon, one D-Bus contract, hybrid local/cloud inference, and an audit log for every byte that goes to the cloud.

View on GitHub How it works

you › what's eating my RAM?

veya › Firefox is using 6.2 GB across 14 processes,
        followed by GNOME Shell at 1.1 GB.
       ↳ local model · 2 MCP tools · 0 bytes sent to cloud

Built around three commitments

🔒

Local-first privacy

Local models are first-class. Every source of context — clipboard, files, notifications, screen — is gated by per-source permissions, default-deny.

👁️

Transparency

An audit log records every tool invocation and every byte sent to a cloud backend. Cloud usage is always user-visible, surfaced live over D-Bus.

🧩

System-wide, not app-bound

One daemon, one D-Bus contract (org.veya.Veya1), any number of frontends — overlay window, shell extension, CLI, or your own client.

Architecture at a glance

A central daemon owns all intelligence. Frontends are thin clients over D-Bus; system access is isolated in an MCP server behind a safety layer.

Daemon

Long-running user service. Owns sessions, context, permissions, the audit log, and the model router. Local-first FallbackInferenceBackend: try local Ollama, fall back to cloud.

McpServer

System-action layer over the MCP protocol. Mostly read-only tools — system info, processes, memory/disk, journald, APT, systemd — plus a permission-gated clipboard write, all through a central safety layer.

Personal context index

Grounds answers in user-approved content via local embeddings + SQLite/sqlite-vec. Permission-checked at both ingestion and query time.

Safety layer

Every shell command goes through an allowlist, timeouts, output caps, and the audit log. Daemon and MCP server run unprivileged.

Roadmap

Pre-alpha. Milestone 1 is the walking skeleton; everything builds on the same daemon + D-Bus + MCP spine.

M1

MVP — end-to-end answer on screen

Daemon skeleton, D-Bus Ask, MCP server with read-only system tools plus a permission-gated clipboard write, Claude API backend, minimal overlay window.
M2

Local models + first write actions

Ollama/LLamaSharp local backend, config-selectable cloud tier, and the first permission-gated write tools (clipboard).
M3

Personal context + notifications

Embeddings index over approved sources; notification intelligence — summarize, prioritize, answer.
M4

Voice, screen awareness, GNOME polish

Voice I/O, permission-gated screen awareness, shell extension shim and keyboard summon.

Open source. Apache-2.0. Contributions welcome.

One issue per branch; every PR runs the canonical build + test + format check in CI.

Star it on GitHub