A self-hosted AI agent platform that turns a kanban board into an autonomous development team. Describe tasks, the operator picks them up, writes code, and delivers results for your review — all on your own hardware, with your own models, and zero data leaving your server. Free for personal use.
- Kanban-driven agents — create tasks, the operator picks them up and works autonomously
- Human-in-the-loop — every result lands in a review column before it's accepted
- Assistant / brain-dump chat — describe an idea in plain language, the system creates structured tasks
- Scheduled triggers — cron-based agent execution without manual intervention
- VS Code integration — MCP server exposes tasks and review directly in your editor
- Bring your own model — connect any OpenAI-compatible endpoint (LocalAi, LiteLLM, …) or OpenAI directly
- 100 % self-hosted — your data never leaves your machine
- 0 % telemetry — no tracking, no analytics, no data collection ... at least not in our code 😄
curl -fsSL https://raw.githubusercontent.com/data-ps-gmbh/Agent-Elno/main/get-agent-elno.sh | sudo bashThe interactive installer asks for your LLM provider, model name, and ports — then starts everything. Same command to update. Default login: admin / password.
curl -fsSL https://raw.githubusercontent.com/data-ps-gmbh/Agent-Elno/main/remove-agent-elno.sh | sudo bashRemoves all services, binaries, config, and data. Asks for confirmation before deleting.
- Open your project in VS Code — Copilot connects to the agent via MCP
- Define tasks with Copilot — describe what needs to be done, Copilot hands the task into the system
- Repeat — queue up everything that needs doing
- The operator works autonomously — picks up tasks, writes code, creates feature branches
- Check out the feature branch for review — rebase to latest if needed
- Review the result — approve, or add comments requesting changes
- Tell the operator to merge via comment and remove the review tag to let it continue
The operator picks up tasks based on column and tags:
- Processed: Any column except Backlog and Done, without the tags blocked or review
- Skipped: Tasks in Backlog, Done, or tagged blocked / review
Move a task to Ready (or any active column) to let the operator pick it up. When the operator finishes, it tags or moves the task to review or done — you check the result and either approve or comment.
Away from VS Code? Use the chat to tell your personal agent things on the go — via the web UI or the mobile app (currently in closed beta).
We run a hybrid setup: a local Qwen model for orchestration and chat, OpenAI Foundry models for coding and review (bound directly or via LiteLLM as a unified proxy).
This is our budget setup — optimized for cost, not peak performance. Bigger models (Claude, GPT-4.5, o3) handle larger projects and produce smarter results. The table below is a starting point, not a ceiling.
The model is everything. Agent-Elno is an orchestration layer — it sets the stage, but the model does the actual thinking. A weak model will produce weak results, no matter how well the system is configured. A capable model will surprise you. Choose wisely, and the agent becomes genuinely useful; choose poorly, and you'll spend more time cleaning up than you saved.
| Role | Model | Local/Cloud | Why |
|---|---|---|---|
| Operator (task orchestration) | Qwen3-32B-Q4_K_M | Local | Better instruction following than Qwen2.5, reliable function calling, free |
| Personal Agent (chat) | Qwen3-32B-Q4_K_M | Local | Great personality, fast responses, keeps data private |
| Document Editor | Qwen3-32B-Q4_K_M | Local | Solid markdown and prose generation |
| Developer (coding) | gpt-5.1-codex-mini | Cloud (OpenAI) | Great quality-to-cost ratio for large projects |
| Senior Developer (complex coding) | gpt-5.1-codex-max | Cloud (OpenAI) | Large context window, handles multi-file changes well |
| Architect (design/planning) | o4-mini | Cloud (OpenAI) | Strong reasoning, good at structural decisions |
| Reviewer (code review) | o4-mini | Cloud (OpenAI) | Thorough review, follows coding guidelines |
| Embedding | nomic-embed-text-v1.5 | Local (LocalAI) | 768-dim, fast, good semantic search quality |
| Model | Role tested | Verdict | Notes |
|---|---|---|---|
| Qwen3-8B | Agent / Operator | ❌ | Bad instruction following |
| Phi-4-Mini-Reasoning | Agent / Operator | ❌ | Hallucinations |
| Phi-4-Mini-Instruct | Agent / Operator | ❌ | Bad function calling |
| Qwen2.5-Coder-32B-Instruct | Coder | ❌ | Poor understanding of large codebases |
| Qwen2.5-Coder-14B-Instruct | Coder | OK for small projects, not for bigger solutions | |
| Qwen3-Coder-30B-A3B | Coder | ❌ | Very fast, very unreliable output |
| Qwen3-Coder-Next | Coder | ❌ | Tends to review instead of code, even with explicit instructions |
| NousResearch Hermes-4-14B | Agent / Operator | Great personality, OK instruction following, bad as operator | |
| Microsoft NextCoder-32B | Coder | Ignores guidelines, OK for smaller projects only | |
| DeepSeek-V3.2 | Coder | Decent output, problematic function calling behavior | |
| GPT-4o | Coder | Decent, replaced by gpt-4.1 | |
| GPT-4.1 | Coder | ✅ | Good context understanding, reliable output — replaced by codex-mini for cost |
| GPT-4.1-Nano | Agent / Operator | Decent reasoning for a small model | |
| Qwen2.5-32B-Instruct | Agent / Operator / Chat | Good instruction following, great personality — but loops on tool calls under load | |
| Qwen3-32B (Q8) | Agent / Operator | Similar looping issues as Qwen2.5, needs R/W/E guard |
For coding tasks:
- gpt-5.1-codex-max (cloud) — very big context for massive coding tasks, redesign or rewrites
- gpt-5.1-codex-mini (cloud) — best value: high quality, large context, low cost
- gpt-4.1 (cloud) — reliable fallback with good context understanding
- Local models struggled with our production codebases (50k–100k lines of C# and 20k–40k lines of Razor per project)
For orchestration / chat:
- Qwen3-32B-Q4_K_M (local) — our pick: better instruction following than Qwen2.5, great personality, zero cost
- Qwen2.5-32B-Instruct still works but tends to loop on tool calls under load
- Smaller models (8B–14B) were unreliable for function calling
For review / architecture:
- o4-mini (cloud) — strong reasoning, follows guidelines well
For embeddings:
- nomic-embed-text — fast, runs locally, good semantic search quality
- Local (LocalAI / Ollama): Free, private, no rate limits — but needs GPU for acceptable speed (32B needs ~24 GB VRAM)
- Cloud (OpenAI / Azure): Faster, smarter coding models — but costs money and data leaves your server
- Hybrid (LiteLLM): Route orchestration locally, coding to cloud — best of both worlds
| Quick Start | Installation and first steps |
| Configuration | Environment files, config modes, all options |
| Architecture | Service architecture and data flow |
| Operator Process | How the autonomous loop works |
| Agents & Skills | Agent definitions and prompt templates |
| Chat & Memory | Personal agent, sessions, semantic memory |
| Scheduler | Cron-based triggers |
| Integrations | LiteLLM, Ollama, OpenAI, nginx, Traefik |
| Troubleshooting | Logs, action log, common issues |
| Changelog | Release history |
- Debian 12+ or Ubuntu 22.04+ (x64)
- An OpenAI-compatible LLM endpoint (Ollama, LiteLLM, OpenAI, …)
- 2 GB RAM, 4 GB disk minimum
No Docker, no .NET SDK, no runtime installation required.
We built Agent-Elno to fit our own workflow — but we're actively developing it further. If you have suggestions, feature requests, or run into problems, we'd love to hear from you.
- Feature ideas — open an issue; if it fits our roadmap, we'll try to integrate it
- Bug reports — please include a brief description of the problem and steps to reproduce; we'll investigate anything we can reproduce ourselves
- Questions — we try to answer as fast as we can
We can't promise everything, but we read every issue and do our best to help.
- GitHub Issues — bug reports, feature requests, questions
- data-ps.de — company website
- info@data-ps.de — commercial inquiries
Free for personal, non-commercial use under the PolyForm Strict License 1.0.0.
Commercial use requires a separate license — contact info@data-ps.de.
© 2014–2026 DATA-PS GmbH. All rights reserved.