Note
Ecosystem & Machine Discovery Index: For machine-readable context, agent-context loading, and comprehensive repository routing, see llms.txt. All active software projects in the ellmos-ai organization operate under local-first principles with SQLite persistence, minimal external dependencies, and transparent component composability.
ellmos (XLLM-OS) is a family of text-based operating systems that empower Large Language Models to work autonomously, learn, and self-organize.
Tip
🗺️ Interactive maps — ellmos-ai.github.io: explore the ecosystem visually. The module circuit map shows every functional area and how they work together, the skill library lets you read and copy every public skill, and the stack composer assembles your own stack with live composition-rule checks.
This index is complete for the public ellmos-ai repositories (75 active repos, 1 archived; 76 total). Archived repositories are marked explicitly. Last checked against GitHub: 2026-09-27.
| Area | Repositories |
|---|---|
| Organization profile | .github - org profile, community health files and llms.txt; ellmos-ai.github.io - interactive maps and ecosystem explorer |
| Stack catalog & recipes | stacks - catalog and shared manifest schema for every stack in the ellmos-ai family; bundles - recipe layer of the ellmos ecosystem (bundle manifests, catalogs and composition knowledge) |
| Operating systems | bach, rinnsal, ellmos, open-ocean - plus gardener as the minimal OS tier when run standalone; it is indexed under Memory and Control below |
| Memory pillar | usmc, gardener, task-master - curated session memory, organic cross-source index, and task tracking; see Memory and Control |
| MCP servers | ellmos-codecommander-mcp, ellmos-filecommander-mcp, ellmos-clatcher-mcp, n8n-manager-mcp, ellmos-controlcenter-mcp, ellmos-homebase-mcp, ellmos-servercommander-mcp, ellmos-blender-use-mcp, open-compute-mcp |
| Agent modules and orchestration | clutch, connectors, MarbleRun, swarm-ai, n8n-workflow-manager, ellmos-stack, agent-ops-stack, skills, build-your-users-mind, open-compute, web-scraper - standalone web scraper (get/links/forms/headers/extract/screenshot) extracted from BACH, with an SSRF guard; anonymizer - local-first document pseudonymization with fail-closed NER; report-forge - domain-neutral core for anonymizable report pipelines |
| Agent hooks, evidence and coordination | memoryhooker - connects local memory sources to coding-agent lifecycle hooks (no network); workflowhooker - configurable workflow checks at agent lifecycle events, zero dependencies; hook-master - local-first pointer registry and materialization for agent hooks (sibling to policy-registry); roshambo - multi-agent coordinator: serializable leases + outcome memory on CockroachDB; roshambo-starmap - evidence artefact of the multi-vendor swarm run coordinated by roshambo; policy-registry - local-first registry for policies, rules and governance decisions with metadata pointers to canonical sources; source-resolver - staged role-based source resolution ladder for skills; grounding-seed - standalone-bootstrap template for isolated modules/skills, delegates to source-resolver when importable; prompt-evidence-collector - audit-ready capture of LLM interactions with a cryptographic, hash-only evidence store |
| Agent operations tooling | ticket-master - multi-provider ticket router and triage console for CLI coding agents; lock-master - portable multi-agent file-lock system; system-gap-master - serverless cross-machine sync yard; system-auditor - evidence-based system audits across machines with meta-audit bundling; compare-race - same prompt to several models, stopwatch or true race, the starting model judges; prompt-listener - fork-master template for prompt/session analysis with AgentEvent provenance schema; coma - agent lifecycle layer (spawn, file protocol, status polling); companion-for-agy - PTY wrapper that makes agy (Gemini CLI) output readable for automation; decision-clicker - file-based decision chains with audit trail, guarded undo and a local web UI; session-checkpoint - application-neutral session checkpoint storage with integrity checks |
| Agents | hungrycall, ringedingeding, researchcall - telephone agents built on CALL-E; plus agent roles shipped inside ticket-master, task-master, build-your-users-mind, law-checker and ai-media-editor, and as skills (research-agent, dev-soft-agent) in skills; see Agents below |
| Competition entries | hungrycall, ringedingeding, researchcall, roshambo, roshambo-starmap, cowork-protocol, FolderHome, NemoFold, sentinel-fleet, build-your-users-mind, bach - see Competition Entries below |
| Core and system infrastructure | sqlite-transit-sync - local-first SQLite synchronization through verified snapshots and configurable row-level merge policies (Python 3.10+, zero dependencies); ellmos-scheduler - standalone local-first task scheduler and run recorder for modular ellmos stacks; ellmos-voice-io - local-first, LLM-neutral STT, TTS, and wake-word pipeline for ellmos and standalone CLI agents |
| Domain tools | law-checker - source-grounded AI first-look legal assessments for German law (Erstorientierung, no substitute for a lawyer), statute registry and embodiment agents; worksheet-generator - generates structured worksheets for pedagogical and therapeutic use, either from an ICF-aware support goal or from subject, grade and topic, rendered to Markdown/HTML/DOCX; steuer-assistent - offline-first worksheet for German employee income-related expenses (Werbungskosten): records self-categorized receipts and sums them to the cent, entirely locally. It does not assess deductibility and does not file a return - not tax advice; pasta-press - local AI text press via Ollama: style refinement, translation & AI-marker cleanup with honest AI disclosure; doc-services - document extraction to text/Markdown with a learning backend preference and honest fallback, AGPL-free OCR, and content-based privacy screening before handoff to third parties; accounts-core - neutral domain core for the account and bank-balance domain: bank account CRUD, CAMT balance import, and a privacy-safe read-only transit projection with strict IBAN masking for downstream consumers (e.g. OCEAN) |
| Media and content workflows | ai-media-editor - local AI video, audio and podcast editing with local transcription, transcript-based cuts, Hyperframes motion graphics and agent-driven creative edits; clip-storyboard-director - local-first AI storyboard director and scene continuity orchestrator bridging video generation, CDP, and multitrack audio |
| Evaluation, templates and maintenance | ellmos-tests - B/O/E evaluation framework for SKILL.md-based LLM operating systems and agent hubs; project-docs-template - agent-ready project documentation template with START/STATE/TODO/DONE, workflows, lightweight tooling and LLM-friendly project memory; system-explorer - evidence-based topology mapping, capability bounds and communication edge discovery; clirec - human-readable GUI demonstration recordings for CLI and agent workflows; ellmos-unified-gui - unified desktop surface that brings the ellmos modules together in one local GUI |
| Legacy archive | recludos-legacy - archived predecessor to BACH |
Our recommended selection — building blocks that integrate into any ellmos OS or stand on their own. The banners are the links; details in the table below:
| Module | Focus |
|---|---|
| swarm-ai | Parallel LLM coordination |
| build-your-users-mind | Per-user theory of mind: decision avatars built from interaction logs |
| web-scraper | Fetch, extract, structure — standalone scraper with an SSRF guard |
| report-forge | Domain-neutral core for anonymizable report pipelines |
| ai-media-editor | Local AI video, audio and podcast editing with transcript-based cuts |
| clutch | Provider-neutral model routing |
| MarbleRun | Chain orchestration |
| n8n-workflow-manager | Local-first n8n management |
| open-compute | Computer-use core with safety gate |
| connectors | Portable messaging connectors & CLI agent bridge |
| anonymizer | Local-first document pseudonymization with fail-closed NER |
| pasta-press | Local AI text press via Ollama: style refinement, translation & AI-marker cleanup with honest AI disclosure |
| project-docs-template | Agent-ready project documentation scaffold with START/STATE/TODO/DONE |
Bundles declare which module versions belong together: versioned, compatibility-checked dependency trees.
The bundles repository hosts the recipe layer of the ellmos ecosystem — bundle manifests, catalogs, and composition knowledge. Our first bundle manifests are registered and undergoing staged rollout. First public candidate: the .MEMORY pillar set usmc + gardener + task-master.
One idea runs through everything in this ecosystem: modules compose into something new. Pick the building blocks you need, wire them your way, and the result is your own stack — not a fixed product edition.
flowchart TD
FLEET["FLEET — same system instances, grouped across hosts"]
SYS["SYSTEM / OS — governance frame above the stacks"]
STACK["STACK — operable composition with boundaries"]
BUNDLE["BUNDLE — dependency tree of modules"]
MOD["MODULE — standalone building block"]
FLEET --> SYS --> STACK --> BUNDLE --> MOD
| Layer | Definition | Public example |
|---|---|---|
| Module | A building block: one standalone capability, independently versioned and useful on its own. | gardener, clutch, every MCP server |
| Bundle | The dependency tree of modules — declares what belongs together, as a versioned, compatibility-checked set. | The .MEMORY pillar set: usmc + gardener + task-master (Memory and Control) |
| Stack | An operable composition with boundaries — declares how it runs together: data, network, tenants, execution. Size classes: bundle, core, full, os-stack. |
ellmos-stack, agent-ops-stack |
| System / OS | The governance frame above the stacks: policies, identity and lifecycle for one installation or edition. | bach, rinnsal, ellmos |
| Fleet | A multi-host grouping of the same system instances — one system, many machines, kept in step. | roshambo with system-gap-master |
Skills play a role on every layer: they are shipped as pluggable modules, versioned inside bundles, wired into stacks, and surfaced by systems to their agents — one skill library, useful from a single module up to a whole fleet.
Stacks declare composition instead of copying module code — so any composition can be re-wired into something new. An operator "control room", for instance, is not a separate product: it is a stack that wires the existing MCP access surfaces (ControlCenter, ServerCommander, Homebase) into a single operations view. Same modules, new whole.
more than a stack
Some compositions outgrow the stack layer: they are governed systems with their own identity, policies and lifecycle. Three of them form our full systems — the banners are the links:
| System | What it is |
|---|---|
| BACH | The stream that unites everything: the full LLM-OS with 113+ handlers, 1870+ skills, boss agents and GUI. |
| Rinnsal | The trickle: lightweight LLM infrastructure — memory, tasks, connectors, chains, i18n. Zero dependencies. |
| open-ocean | Free the ocean: the free community full system of the ellmos ecosystem. |
Different philosophies, same goal — and gardener doubles as the minimal OS tier when run standalone (see Memory and Control).
Stacks are manifest-driven compositions (ellmos.stack.v2) — no code copies, just declared components. Two active public stacks anchor the family, catalogued in a third:
| Stack | Purpose | Core modules |
|---|---|---|
| stacks | Catalog and shared manifest schema for every stack in the ellmos-ai family | — |
| ellmos-stack | Self-hosted, local-first AI research base: Ollama, n8n, Rinnsal memory, Docker Compose automation | Rinnsal · KnowledgeDigest (file-bricks) · Ollama · n8n |
| agent-ops-stack | Coordination layer for CLI coding agents: ticket routing, file locking, cross-machine sync, decision-avatar, MCP control plane | ticket-master · lock-master · system-gap-master · build-your-users-mind · skills · ellmos-controlcenter-mcp · ellmos-homebase-mcp |
A number of these modules are deliberately both: standalone dev tools you can adopt individually, and stack components you get automatically by installing the stack. That also applies to llm-note (doc-bricks) — local-first notebooks for LLM agents, built as a pluggable module for stack composition.
Nine MCP servers, one control plane — arranged as a vertical family tree (Stammbaum): Root & main trunk at the bottom (earliest servers, 2026-02), branching upwards through mid-tier infrastructure (2026-05 to 2026-06) to the youngest twigs at the top (2026-07) with server logos hanging as fruits on the branches.
| Server | Focus | Install |
|---|---|---|
| CodeCommander | Code analysis & refactoring | npm i -g ellmos-codecommander-mcp |
| FileCommander | File management & batch ops | npm i -g ellmos-filecommander-mcp |
| Clatcher | File repair, format conversion, duplicates | npm i -g ellmos-clatcher-mcp |
| n8n Manager | n8n workflow automation | npm i -g n8n-manager-mcp |
| ControlCenter | MCP profile dashboard, capability bundles & policy audits | npm i -g ellmos-controlcenter-mcp |
| Homebase | Local LLM memory, knowledge, state & orchestration | npm i -g ellmos-homebase-mcp |
| ServerCommander | Server health checks, log analysis, deploy dry-runs | npm i -g ellmos-servercommander-mcp |
| Blender Use | Headless Blender asset QA | npm i -g ellmos-blender-use-mcp |
| open-compute-mcp | Computer use: screenshots, safety-gated actions | npx open-compute-mcp |
The family's memory pillar and its coordination & control modules — first their banners, then the details:
| Module | Role |
|---|---|
| usmc | Curated session/core memory — the facade and entry point of the memory system. Push model: "what I consciously remember." |
| gardener | Memory supplier: organic growth via absorb/decay, plus a federated cross-source FTS5 index via observe() that cites results back to their source. Pull model: "index what's already there." Doubles as the minimal OS tier when run standalone. |
| task-master | Standalone SQLite task module — tasks stay separate from knowledge memory. Zero dependencies. |
| ellmos-scheduler | Standalone local-first task scheduler and run recorder for modular ellmos stacks and autonomous LLM agent execution loops. |
| ticket-master | Cross-platform, multi-provider ticket router / triage console — files structured tickets and routes them to the right AI provider or sub-agent. |
| lock-master | Portable multi-agent file-lock system — LOCK*.txt-based project/component locking with scopes, expiry and stale-cleanup. |
| policy-registry | Local-first registry for policies, rules, and governance decisions with metadata pointers to canonical sources. |
| hook-master | Local-first pointer registry and materialization for agent hooks — sibling to policy-registry. |
| source-resolver | Role-based source resolution for skills: a staged ladder (user config, own module, discovery proposal, foreign provider, dialogue) instead of hard-wiring every information source. |
| grounding-seed | Standalone-bootstrap template for isolated modules/skills/repos — self-knowledge, search, self-provisioning, migration on find; delegates to source-resolver when it is importable. |
| system-gap-master | Serverless sync yard for multi-machine, multi-agent setups — slot rule, gated daily ritual, bootstrap runbook. Family: lock-master, ticket-master. |
| system-auditor | Evidence-based system audits across machines — four audit tokens (time, domain, system, auditor), an aggregation ladder with enforced identifiability, model-manual meta reports. Extracted from ticket-master's TICKET-WRITER. |
| compare-race | Same prompt to several models — sequential (stopwatch, clean per-lane timing) or parallel (true race), with repetitions per model; the starting model judges across quality, correctness, completeness, instruction fidelity and latency (time is only one dimension). Reuses the system-auditor identity logic. |
| prompt-listener | Versioned fork-master template: 5-stage prompt/session analysis pipeline plus a stdlib-only AgentEvent v2 provenance schema (18 sub-ledgers). Fork it when your purpose diverges — the only duty is a FORKS.md entry. |
| coma | COMAS — COMmunication for Autonomous Subagents: lifecycle layer for agents (spawn, file protocol, status polling). Zero dependencies, standard library only. |
| memoryhooker | Connects local memory sources to coding-agent lifecycle hooks (no network). |
| workflowhooker | Configurable workflow checks at agent lifecycle events, zero dependencies. |
| companion-for-agy | PTY-based wrapper that captures agy (Gemini CLI) responses via ANSI color extraction — lets Claude Code, Codex and CI pipelines read Gemini output reliably. |
| decision-clicker | File-based decision chains with audit trail and guarded undo — library, CLI and a local web UI. |
| session-checkpoint | Application-neutral session checkpoint storage with integrity checks — snapshot a run, resume it later. |
| prompt-evidence-collector | Audit-ready capture of LLM interactions with a cryptographic, hash-only evidence store. |
Modules built for one concrete field rather than for the agent runtime itself — the banners are the links, details in the Public Repository Index above:
Telephone agents built on CALL-E — each one takes a spoken task off a person's hands and reports back what actually happened, including when nobody picked up.
| Agent | What it does |
|---|---|
| hungrycall | A sequential call cascade for food delivery, table reservations and pickup: ranks candidate restaurants against the user's boundaries, then calls them one after another and stops at the first success. The generalized cascade pattern is documented separately from the food use case. |
| ringedingeding | Asks one question to several people in your own circle and merges the replies into a single result — either intersecting availability to find a date, or reporting the leading tendency together with countervoices and reasons. It does not turn dissent into a false consensus. |
| researchcall | A standardized telephone survey runner with methodological honesty: builds the instrument from its own gated stations, draws a random sample, calls each person once by default, and reports nonresponse instead of collapsing distinct outcomes. Runs fully local as a dry run by default — no account and no real call needed. |
Beyond the telephone agents, several ellmos-ai modules ship their own agents or agent roles:
| Module | Agent role |
|---|---|
| ticket-master | TICKET-MASTER — a long-lived router/triage agent that files structured tickets and dispatches them to the right AI provider or sub-agent |
| task-master | Three operating roles shipped as agent prompts: TASKSOLVER (works the queue), TASKWRITER (captures tasks), MAINTAINER (keeps the task database healthy) |
| build-your-users-mind | Decision avatar — an agent that learns its user's decision patterns from interaction logs and predicts or takes decisions in their spirit |
| law-checker | Statute embodiment agents — configured statutes (e.g. the German constitution and civil code) "speak" as agents — plus a source-grounded first-look assessment agent |
| ai-media-editor | Agent-driven creative editing — performs transcript-based cuts and motion-graphics passes on local media |
| clip-storyboard-director | Storyboard director & scene continuity agent — coordinates local AI video generation passes, shot composition, camera timing, and multitrack audio sync |
Agents also ship as skills in the skills library: research-agent — research pipeline for PubMed and arXiv with quick search and structured literature reviews, pure Python standard library, extracted from BACH's ResearchAgent — and dev-soft-agent — automated software-development pipeline that scans projects, prioritizes tasks and orchestrates development loops.
Agent infrastructure — coordination, orchestration and lifecycle for agents you bring yourself — lives in swarm-ai, roshambo, MarbleRun and coma.
Projects built for public hackathons and competitions. Listed as entries — no placement claimed. The telephone agents hungrycall, ringedingeding and researchcall are shown under Agents above.
| Entry | Competition | What was submitted |
|---|---|---|
| hungrycall | CALL-E "Your Code Is Calling" (Devpost, 2026) | Sequential calling cascade for food delivery and reservations, plus the generalized cascade pattern as the reusable contribution |
| ringedingeding | CALL-E "Your Code Is Calling" (Devpost, 2026) | Multi-recipient response aggregator: one question, several people, one merged answer |
| researchcall | CALL-E "Your Code Is Calling" (Devpost, 2026) | Standardized telephone survey runner with an eight-station research pipeline and honest nonresponse reporting |
| roshambo | CockroachDB × AWS Hackathon (2026-07) | Multi-agent coordinator: serializable leases and outcome memory on CockroachDB, with an MCP interface |
| roshambo-starmap | CockroachDB × AWS Hackathon (2026-07) | Evidence artefact of the accompanying multi-vendor swarm run — a replayable field record of 27 agents coordinating through roshambo |
| cowork-protocol | The WebMCP Challenge (Devpost, 2026) | Reusable collaboration protocol and adaptive bridge for people, WebMCP apps, and web agents |
| FolderHome | Agents for Humans (Devpost, 2026) | Local-first Strands agent for safe household document workflows |
| NemoFold | Nebius Global AI Hackathon (Devpost, 2026) | Evidence-first document agent with persistent local memory |
| sentinel-fleet | All Things Agentic Hackathon (Devpost, 2026) | Autonomous fleet security and monitoring agent cluster |
| build-your-users-mind | Agent recipe entry | A recipe for any AI agent to build a self-improving theory-of-mind model of its user from interaction logs |
| bach | Agent OS entry | The full local-first LLM operating system: memory, handlers, skills, agents and GUI |
These projects live in sibling organizations but are particularly relevant to the ellmos multi-agent ecosystem:
| Project | Org | Description |
|---|---|---|
| llm-note | doc-bricks | Local-first notes and notebook inboxes for LLM agents — extracted from BACH Notizblock/Denkarium patterns with SQLite, plain-text notebooks and six locales |
| knowledgedigest | file-bricks | Local-first knowledge base with LLM preprocessing — ingest, structure and query documents without cloud dependencies; core module of ellmos-stack |
ellmos-ai operates as the primary AI infrastructure and LLM operating system engineering pillar within an integrated open-source software and research ecosystem:
| Organization | Domain Focus | Key Repositories / Role |
|---|---|---|
| open-bricks | Umbrella / Dach-Organisation | Central catalog, open-source software umbrella, ecosystem index |
| ellmos-ai | AI Agent Infrastructure | bach, rinnsal, open-ocean, skills, agent-ops-stack, clip-storyboard-director |
| file-bricks | Desktop Data Tools | ProFiler, ExplorerPro, ProSync, AmpelClip, knowledgedigest, WinStorePackager |
| doc-bricks | Document & Media Systems | DokuReader, MediaBrain, UniversalInvoiceMail, CleanMarkdown, llm-note |
| dev-bricks | Developer & Code Tools | DevCenter, ApiProber, app-rotator, CodeBox, WikiStub-Seed |
| research-line | Open Science & Math Physics | functional-stability-theory, crm-cosmology, fst-nash, rh-even-dominance |
| biotec-line | Bioinformatics & Genomics | VFDistiller, genotype-to-vcf |
| entertain-and-more | Games, RPG & Podcast Tools | ChatAndChess, rpx, KlangpultLight |
| assistassets-ai | Local Financial Analytics | FinancialProof |
| um-bruch | Applied Health & Public Policy | Clinical risk analysis, prescribing models, systems medicine |
| lukisch | Personal GitHub Profile | Developer portal & flagship showcases |
ellmos-ai
local-first AI agent infrastructure
BACH LLM operating system Python SQLite
Rinnsal lightweight LLM memory
clip-storyboard-director AI storyboard director scene continuity orchestrator
ai-media-editor local AI media video audio editing
Model Context Protocol MCP servers
open-ocean free community full system
roshambo CockroachDB serializable lease multi-agent coordinator
ticket-master lock-master agent operations tooling
curated session memory organic cross-source FTS5 index
 recludOS Archived predecessor to BACH Historical reference |
Full documentation | License: MIT | 🇩🇪 Deutsche Version
