From a spring to a stream -- LLM operating systems that flow.
ellmos (XLLM-OS) is a local-first family of text-based operating systems for Large Language Models: agent memory, task state, MCP tools, skills, connector bridges, model routing, and multi-agent orchestration in small SQLite-backed projects. It is designed for people who want an AI operating layer they can inspect, run locally, and extend without a hosted platform.
Quick links: Organization | BACH | Rinnsal | gardener | llms.txt
| If you are looking for... | Start with | Why |
|---|---|---|
| A full personal LLM operating system with GUI, skills, scheduler, bridges, and multi-agent workflows | BACH | Largest ellmos system and the main integration point |
| A lightweight Python infrastructure layer for memory, tasks, connectors, and chains | Rinnsal | Small, local-first, dependency-light foundation |
| A minimal LLM-native SQLite experiment | gardener | One-table operating substrate for simple agents |
| Shared memory, model routing, chains, or parallel agent patterns | USMC, clutch, MarbleRun, swarm-ai | Standalone modules that can be combined with any OS tier |
| MCP servers for Claude Code, Cursor, or other AI IDEs | CodeCommander, FileCommander, n8n Manager, ControlCenter, Homebase, ServerCommander | Tool servers for code, files, workflows, local state, and local control planes |
| Question | BACH | Rinnsal | gardener |
|---|---|---|---|
| I want a full-featured agent OS with GUI, skills, and multi-agent orchestration | Yes | ||
| I want lightweight LLM infrastructure with zero dependencies | Yes | ||
| I want the simplest possible LLM-native OS (1 table, 4 functions) | Yes | ||
| I need Telegram/Email/WhatsApp connectors | Yes | Yes | Planned |
| I want to self-extend with new skills at runtime | Yes | ||
| I want minimal footprint (~2k lines) | Yes | Yes |
The full LLM operating system. 109+ handlers, 373+ tools, 932+ skills, 5 boss agents with 28 experts, PySide6 desktop GUI, scheduler, bridge system, and self-extension via bach skills create.
git clone https://github.com/ellmos-ai/bach.git
cd bach && pip install -r requirements.txt
python system/setup.py
python bach.py --startupLightweight LLM infrastructure: memory, tasks, connectors, chains. Zero external dependencies. Everything BACH does conceptually, but in ~2,000 lines for developers who want to build their own agent on top.
git clone https://github.com/ellmos-ai/rinnsal.git
cd rinnsal && pip install -r requirements.txtLLM-native OS: 1 SQLite table (everything), 4 functions, FTS5 full-text search. Everything is searchable. The LLM is the agent -- gardener just provides the soil.
git clone https://github.com/ellmos-ai/gardener.git
cd gardener && pip install -r requirements.txt+-------------------------------------------------+
| Choose Your OS Layer |
| |
| BACH (full) Rinnsal (light) gardener (min) |
| +---------+ +------------+ +----------+ |
| | 932 | | Zero deps | | 1 table | |
| | skills | | Connectors | | 4 funcs | |
| | 5 boss | | Chains | | FTS5 | |
| | agents | | Events | | = search | |
| +----+----+ +-----+------+ +-----+----+ |
| +---------------+----------------+ |
| | |
| +---------------+---------------+ |
| | Pluggable Modules | |
| | | |
| | USMC -- shared memory | |
| | clutch -- model routing | |
| | MarbleRun -- agent chains | |
| | swarm-ai -- parallel LLMs | |
| +-------------------------------+ |
+-------------------------------------------------+
| BACH | Rinnsal | gardener | |
|---|---|---|---|
| Philosophy | Maximalist: everything integrated | Lightweight: zero dependencies | Minimalist: 1 table, 4 functions |
| Database | SQLite (145+ tables) | SQLite (structured) | SQLite (1 table everything + FTS5) |
| Memory | 5-type cognitive model | Facts/Notes/Lessons/Sessions | Unified (memo/lesson/recall + decay) |
| Tasks | Full GTD (priority, deadline, tags) | Priority + Status + Agent assignment | type='task' in everything |
| Tools | 373+ specialized tools | CLI commands | 6 bridge+skin tools (extensible) |
| Skills/Agents | 932 skills, 5 boss agents, 28 experts | None | None (the LLM is the agent) |
| Connectors | Telegram, Email, WhatsApp | Telegram, Discord, Home Assistant | Planned (v0.2+) |
| GUI | PySide6 Desktop + Web | CLI only | CLI only |
| Self-Extension | bach skills create |
No | No |
| Codebase | ~50,000+ lines | ~2,000 lines | ~1,600 lines |
| Best for | Power users, all-in-one | Developers wanting light infra | Minimalists, LLM-native experiments |
These modules integrate into any ellmos OS -- or work standalone:
| Module | Purpose | Key Feature | Repo |
|---|---|---|---|
| USMC | Cross-agent shared memory | Confidence-based conflict resolution, change tracking | ellmos-ai/usmc |
| clutch | Provider-neutral model routing | Auto-learning which model fits which task, budget zones | ellmos-ai/clutch |
| MarbleRun | Chain orchestration | Autonomous multi-round agent loops with context handoff | ellmos-ai/MarbleRun |
| swarm-ai | Parallel LLM coordination | 5 patterns: Epstein, Hierarchy, Stigmergy, Consensus, Specialist | ellmos-ai/swarm-ai |
gardener has two roles, and both are current. Run standalone, it is the minimal OS tier described above: one table, four functions, FTS5 search. Composed into the .MEMORY pillar, it is the organic cross-source index alongside usmc (the curated façade) and taskplan (task state). Which role applies depends on how you deploy it — see the organization profile for the pillar view.
ellmos provides Model Context Protocol servers for integration with Claude Code, Cursor, and other AI-powered IDEs:
| Server | Tools | Description | Install |
|---|---|---|---|
| CodeCommander | 18 | Code analysis, refactoring, import management, JSON/encoding repair | npm i -g ellmos-codecommander-mcp |
| FileCommander | 45 | File management, batch operations, process control, async search, cloud-lock checks | npm i -g ellmos-filecommander-mcp |
| Clatcher | 12 | File repair, format conversion, duplicate detection, batch operations | npm i -g ellmos-clatcher-mcp |
| n8n Manager | 18 | Create, update, back up, and manage n8n workflows | npm i -g n8n-manager-mcp |
| ControlCenter | 20 | Alpha control plane for local MCP servers, Claude profiles, policy audits | npm i -g ellmos-controlcenter-mcp |
| Homebase | 45 | Alpha MCP server for local LLM memory, knowledge, state, routing, testing, and orchestration | See repo README |
| ServerCommander | 8 | Alpha MCP server for deploy dry-runs, mail status, log analysis, and server health checks | See repo README |
| Blender-Use | 3 | Headless Blender asset QA: background render, FBX re-import verification, script execution | npm i -g ellmos-blender-use-mcp |
| open-compute | 10 | Computer-use over MCP: screen capture, semantic UI targeting via Windows UIA, canonical actions | npm i -g open-compute-mcp |
| Project | Description | Repo |
|---|---|---|
| skills | Pluggable skill library (dev, research, education, infrastructure) | ellmos-ai/skills |
| n8n Workflow Manager | Standalone GUI for n8n workflow creation | ellmos-ai/n8n-workflow-manager |
| ellmos-stack | Self-hosted AI stack (Docker, Ollama, n8n, memory, knowledge base) | ellmos-ai/ellmos-stack |
| ellmos-tests | Cross-OS test suite and benchmark reports | ellmos-ai/ellmos-tests |
| stacks | Catalog and shared manifest schema for every stack in the family | ellmos-ai/stacks |
| agent-ops-stack | Manifest-driven composition of the local agent-ops ecosystem | ellmos-ai/agent-ops-stack |
| connectors | Portable messaging connectors: Telegram, Discord, Signal, WhatsApp, Home Assistant, Webhook | ellmos-ai/connectors |
| open-compute | Model-agnostic computer-use core for Claude, OpenAI CUA and mock backends | ellmos-ai/open-compute |
| web-scraper | Standalone web scraper with SSRF guard, extracted from BACH | ellmos-ai/web-scraper |
| taskplan | Deterministic task selection so the backlog cannot hide | ellmos-ai/taskplan |
| anonymizer | Local-first document pseudonymization with fail-closed NER | ellmos-ai/anonymizer |
| report-forge | Domain-neutral core for anonymizable report pipelines | ellmos-ai/report-forge |
| project-docs-template | Agent-ready project documentation template (START/STATE/TODO/DONE) | ellmos-ai/project-docs-template |
| clirec | Human-readable GUI demonstration recordings for CLI and agent workflows | ellmos-ai/clirec |
| ai-media-editor | Local AI editing for video, audio and podcasts with local transcription | ellmos-ai/ai-media-editor |
| law-checker | Source-grounded first-look legal assessments for German law | ellmos-ai/law-checker |
| steuer-assistent | Offline-first worksheet for German employee income-related expenses | ellmos-ai/steuer-assistent |
| worksheet-generator | ICF-aware worksheets for pedagogical and therapeutic use | ellmos-ai/worksheet-generator |
| build-your-users-mind | Per-user theory of mind: a decision avatar built from interaction logs | ellmos-ai/build-your-users-mind |
- Pick your OS tier using the comparison table above
- Clone and install using the quick-start commands
- Optionally add modules (USMC for shared memory, clutch for model routing, etc.)
- Add MCP servers for IDE integration:
npm i -g ellmos-codecommander-mcp ellmos-filecommander-mcp
All projects: Python 3.10+ | SQLite | MIT License | Zero or minimal dependencies
ellmos is the ellmos-ai local-first LLM operating-system family. It is not Eclipse LMOS, AllenAI OLMo, ELMo embeddings, Elmo Software, or a hosted agent platform. Useful search phrases include:
ellmos-ai ellmos local-first LLM operating systemellmos BACH Rinnsal gardener SQLite agent OSellmos MCP CodeCommander FileCommander n8n Managerlocal-first LLM OS SQLite memory skills MCPClaude Code MCP local filesystem code analysis ellmosellmos-ai ellmos canonical LLM OS namespace
The canonical GitHub namespace is ellmos-ai. Older search-index snippets may still show lukisch/ellmos or legacy bach-* package names; use the ellmos-ai/* repositories and current ellmos-* MCP package names for new installs.
For automated indexing and AI assistants, see llms.txt.
All ellmos projects are released under the MIT License.
- Organization: github.com/ellmos-ai
- Author: Lukas Geiger
ellmos -- Extra Large Language Model Operating Systems The stream that unites everything.