Why I built Roger
WHY I BUILT ROGER - SELF-HOSTED AI AGENTS FOR DEFENSE
There's a shift happening in AI right now. Anthropic just shipped Dispatch and Computer Use for Claude - you can text your agent from your phone, it takes over your desktop, opens apps, navigates screens, and delivers finished work while you're away. Not a summary to review. Not a draft to edit. The work itself, done.
This is what people have been doing with OpenClaw for months. And now that Anthropic is shipping the managed version, the obvious question is: why would anyone self-host?
I'll tell you why. Because I've spent 40 years building software for defense and federal clients, and managed platforms aren't an option when your data has to stay on your hardware.
THE PROBLEM
Regulated industries, air-gapped networks, defense systems, data sovereignty requirements - there's a large world of work where "just use Claude's managed stack" isn't the answer. You need to own the infrastructure. You need to audit the binary. You need to control which models your agents can reach and which data they can touch.
That world didn't have a good option for AI agents.
WHAT ROGER DOES
Roger is a self-hosted AI agent gateway. Single Go binary, 15MB, runs as a systemd service. You give it a task from Telegram, Discord, Slack, or HTTP, and it handles it.
It routes to any model provider - Anthropic, OpenAI, GitHub Copilot, Ollama, OpenRouter. If one goes down, it falls through to the next. You're not locked to any vendor.
It runs a cron scheduler. You set up overnight jobs. You wake up to finished work.
It does multi-agent orchestration. Complex tasks get split into parallel subtasks with dependency ordering. Your agents coordinate without you managing each step.
It automates browsers via chromedp - for the apps that will never have an API. Old Jira instances, bespoke ERP screens, antique SAP systems. The kind of work that eats half your day because you're manually copying data between screens.
And it ships an MCP server that plugs into Claude Code, Claude Desktop, and Cursor - so you get the best of both worlds.
THE COMPANION: EIDETIC
The other half of the stack is Eidetic, a semantic memory server. Most AI agents forget everything between conversations. Eidetic fixes that.
Built on PostgreSQL + pgvector + Ollama, it stores memories, embeds them for semantic search, and makes them available to any agent via MCP. Self-hosted, CPU-friendly, no cloud required.
When you pair Roger with Eidetic, your agents don't just execute tasks - they build context over time. They notice patterns across weeks. They connect a hiring trend to a strategic conversation from last month. This is what Nate B Jones calls "compound signal detection," and it's the real unlock.
SELF-HOSTED VS MANAGED
Nate put it well: the difference isn't safety. It's the degree to which you want to self-host versus the degree to which you want a managed agent.
Anthropic's managed stack is great for most people. But if you need model independence, data sovereignty, or the ability to deploy in environments that will never touch a commercial cloud API - that's Roger's lane. And for teams operating in GovCloud, IL4/IL5, or air-gapped enclaves where even FedRAMP-authorized SaaS doesn't cut it, self-hosted is the only option.
Every infrastructure shift follows this pattern. Self-hosted email to Gmail. Rack servers to AWS. Jenkins to GitHub Actions. The self-hosted version proves the category exists. The managed version gets mass adoption. Both have their place.
WHAT'S NEXT
Roger and Eidetic are both MIT-licensed and actively maintained. If you're in defense, federal, or regulated work and you've been waiting for an AI agent stack you can actually deploy - this is it.
GitHub: https://github.com/EMSERO/roger GitHub: https://github.com/EMSERO/eidetic
Or reach out at info@emsero.com.
- Brian Brown, EMSERO
