Your sovereign
local AI harness
Say goodbye to US cloud datacenters for confidential documents, code, and company secrets. Run open-source LLMs deterministically, resource-efficiently, and auditably on your own infrastructure — with full workflow graphs, isolated namespaces, and Human-in-the-Loop.
gemma4:e2b via Ollama starting at 6 GB VRAM) – no GPU clusters needed.
End Context Bloat & Prompt Roulette
Why traditional agent frameworks struggle on local hardware and how Libre Workspace AI delivers deterministic precision.
Too context-heavy & erratic
Traditional frameworks cram the entire chat history, endless system prompts, and dozens of tool definitions into a single massive prompt context.
-
Context Degradation: Local models lose track and hallucinate when prompts contain thousands of irrelevant tokens.
-
VRAM Spikes: High compute waste and sluggish response times on standard desktop and server hardware.
-
Infinite Loops: Uncontrolled agent loops without strict barriers or deterministic step validation.
Deterministic, lean & focused
Execution flow, state transfers, and guardrails run in deterministic Python code. The LLM is called strictly for language comprehension and synthesis.
-
Dynamic Skill Runner: Injects only the exact tool needed for the current step at runtime into the prompt.
-
Strict Loop Guards: Automatic safety guard stops execution after 20 iterations per run to prevent runaway processes.
-
Transparent File System:
SOUL.md,memory/, andlearnings/stored in open, versionable Markdown files.
For Linux Desktop
Installs unprivileged with a single command in user space (~/.local/share/libre-workspace-ai). Includes a systemd user service and native desktop app launcher – ready immediately without complex server setups.
For Server Operators & Enterprises (OIDC)
Enterprise multi-tenancy with OpenID Connect (OIDC) Single Sign-On (e.g. Keycloak, Authentik). Automatic synchronization of security groups into USER.md and role-isolated agent sandboxes.
Three Pillars of True Sovereignty
From interactive multi-modal chat and relentless business automation graphs to modular tool extensions.
Flexible AI Chat & Dynamic Skill Runner
Use a familiar, responsive chat interface for spontaneous inquiries, brainstorming, and ad-hoc tasks. Under the hood, the intelligent Dynamic Skill Runner ensures maximum efficiency.
-
✓
Context Protection: Feeds the LLM only the tools needed for the immediate step, preserving token budget.
-
✓
No Overload: Lean models (from 6 GB VRAM) stay focused with significantly fewer hallucinations.
-
✓
Multi-Modal: Built-in offline voice recognition (Vosk), document processing, and vision analysis.
Deterministic Workflows (n8n-Style)
Fixed, directed execution graphs for mission-critical business processes. Define workflows in Python or using the visual Canvas Builder for 100% reproducibility.
-
✓
Directed Graphs: Explicit chaining of
BaseNode,PythonDecisionNode, and LLM nodes. -
✓
Automated Pipelines: Daily email digests, system health checks, automated document generation (PDF/ODT).
-
✓
Auditable & Reliable: No prompt gambling — every single transition and output is tracked and reproducible.
Modular Skills & MCP Integration
Extend capabilities seamlessly across three standardized skill interfaces tailored for your infrastructure.
-
✓
Pure Python Domain Skills: Direct access to IMAP/SMTP, Nextcloud WebDAV, Matrix, Bash, System Ops, and ODT/PDF creation.
-
✓
Markdown Prompt Skills: Define specialized agent behaviors with simple
SKILL.mdfiles. -
✓
Model Context Protocol (MCP): Native support for community MCP servers and external tool ecosystems.
Python Skills
Native Python execution for IMAP, WebDAV, Bash, System & ODT/PDF
Prompt Skills
SKILL.md templates, roles & reusable agent heuristics
MCP Protocol
Standardized Model Context Protocol for 3rd-party tools & MCP servers
Sovereignty Over Autonomy
Fully autonomous agents carry serious real-world risks. Libre Workspace AI enforces human approval gates and strict task sandboxing.
Human-in-the-Loop
HumanApprovalNode • Approval Gate
Any critical node execution pauses the pipeline and issues an interactive approval card in the WebGUI and CLI. Inspect outputs, edit parameters inline (
✏️), and authorize via /approve — or reject safely.
apt update && apt upgrade -yRisk Level: Medium (Root privileges)
Agent Namespaces
Multi-Agent • Attention IsolationLibre Workspace AI separates operational domains (Sysadmin, Support, Copywriter, Analyst) into isolated namespaces with dedicated long-term memory, personality profiles (
SOUL.md), and self-reflection logs.
├── klaus_martin/ (Department Lead & Router)
│ ├── SOUL.md • memory/ • learnings/
├── operator/ (Automation & Operations)
├── copy_pr/ (Ghostwriter & Communications)
└── librarian/ (Knowledge Curator & Synthesis)
Installation
Choose the best option for your desktop, self-hosted server, existing Libre Workspace setup, or managed GDPR cloud.
Desktop 1-Click Installer
Installs Libre Workspace AI into unprivileged user space (~/.local/share/libre-workspace-ai) including Ollama, systemd user service, and desktop application launcher.
Standalone Docker Deployment
Containerized deployment with Docker Compose. Includes enterprise multi-tenancy, OIDC single sign-on, and configurable AI inference backends.
Libre Workspace Add-on
Simply search for lw-ai in the Addon Center. Installs and links the AI components automatically into your desktop workspace.
Managed Libre Workspace Cloud
Hosted in ISO-27001 certified data centers (Germany & Finland), 100% GDPR-compliant, pre-connected to Nextcloud, Email, and Matrix.
Shape the Future of Libre Workspace AI!
Have ideas for new skills, feedback on architecture, or want to share your custom workflows? Connect with our community or send us your feedback directly.