Everyone claims it. No one ships it. The open-source core is real — and free.
There is a familiar shape to the modern AI-tooling launch. A gorgeous landing page. A bold one-liner about agents that "do real work." A looping hero video. And then, where the product should be: Join the waitlist. Request access. Book a demo. Coming soon.
For a long while that was the whole category. It isn't anymore — so this is the honest version of where things actually stand, competitors named, with the one thing that's still genuinely different called out plainly.
The category caught up — mostly
Credit where it's due: there is now a real, well-funded class of agentic AIOps / "AI SRE" products, and several of them ship working software. Komodor turned its Klaudia engine into an extensible multi-agent platform — dozens of specialized agents spanning Kubernetes, GPUs, networking and storage — and showed it at KubeCon EU 2026. Resolve.ai reportedly raised over $100M at a billion-dollar valuation in early 2026 around an "AI for production" agent. Cleric runs hypothesis-driven investigations and landed on Gartner's Cool Vendor list. NeuBird, Causely, and others are in the same lane. These are not waitlists; they resolve real alerts for real customers, and pretending otherwise would be dishonest.
So the honest headline isn't "nobody ships." It's narrower and more interesting: almost everyone ships the same shape — a SaaS you rent, running in their cloud, on a model they chose (or a short "approved" list), reaching into your systems as a service account, with your telemetry flowing to their servers. Traversal is the exception that proves the rule — it offers on-prem deployment and lets customers bring pre-approved models, which is exactly the direction this whole post is about. Good. The rest of the differences are about how far past that you can actually go.
The part that's still different: it runs as you
Here's the distinction I went looking to disprove and couldn't. Ask how any of these agents actually touch your cloud, and the answer is almost always the same: a dedicated agent identity — a service account or managed identity — with read-only RBAC, a command allowlist, and a human approval gate on anything that mutates. Even Microsoft's Azure SRE Agent works this way: it acts as its own managed identity, and "on-behalf-of" the human is an escalation exception an administrator has to authorize, not the default. The agent is a robot with its own badge. When it does something, the audit log says the robot did it.
Our open-source core inverts that. Identity is anchored in your IdP (Entra / Azure AD), and through on-behalf-of token exchange the agent federates the human's identity into AWS (IAM Identity Center / STS), GCP (Workload Identity Federation), and Kubernetes (OIDC → RBAC) — so the action runs as the person, scoped to exactly what that person is allowed to do, and CloudTrail, Cloud Audit Logs, and the Kubernetes audit log all record a name, not a shared bot. Your cloud provider stays the primary authority on who can do what; the agent never becomes a standing key. The full architecture write-up is here.
I went looking for another platform that combined all of it — run-as-the-user OBO identity, agentic tools, no-code Flows, and Code Mode, self-hosted, on any model or provider. I didn't find one. That is not the same as claiming none exists — the space moves weekly and I haven't read every datasheet — but I can tell you what I couldn't find, and I'd rather say it that way than fake the certainty.
Run it anywhere on the spectrum — your call, not ours
The other thing the rented-SaaS shape can't give you is a real choice about where it runs. Our open-source core is built to deploy across the whole spectrum, and the operator picks the point:
- Fully air-gapped. On-prem GPU clusters running Ollama (and your own fine-tuned models), zero external LLM calls, nothing leaving the network. The hard mode most of the category simply can't do.
- Hybrid — any permutation. An air-gapped Ollama layer for sensitive work plus private-link / same-VNet / interconnect paths to managed model endpoints — Vertex, Bedrock, Azure AI Foundry — when you want frontier capability without touching the public internet. Route per capability: local for one job, a CSP endpoint for another.
- SaaS / managed. For teams that would rather we run it, our hosted enterprise tier is the same platform — the easy button, with the honest trade-off that you hand back some of the autonomy the self-hosted modes give you.
That flexibility isn't a checkbox; it's the reason the architecture exists. It's built for the environments where "just send it to our cloud" is a non-starter — regulated finance, defense, healthcare, the public sector — and where every action has to trace to a person. The same architecture is designed to meet FedRAMP High and NIST 800-53 high-impact controls: a design target the platform is built to satisfy, not a certification we claim to already hold.
Claim vs. ship
Set against the typical launch — and most of the rented-SaaS field — the difference is concrete:
| The typical "AI agent" / AI-SRE launch | The open-source core | |
|---|---|---|
| Where it runs | Their cloud (a few offer on-prem) | Your laptop, your cluster, fully air-gapped — your choice |
| The model | Theirs, or a short "approved" list | Any — Ollama local, Vertex, OpenAI, Azure, Bedrock; swap per chat, per capability |
| Identity | A shared bot / service account, read-only + approval gate | Runs as the human, on your IdP — every action attributed to a person |
| Your data & telemetry | Flows through their servers | Never leaves your infrastructure |
| Getting started | Waitlist · sign-up · "coming soon" | curl -sSL https://install.openagentics.io | bash |
| Price | Per-seat, per-token, per-investigation | Free to self-host. You pay your provider, nobody else. |
| Audit | Their dashboard, their retention | Tamper-evident, checksummed action log, on your infra |
What's actually real today — no asterisks we're hiding
We'd rather under-claim and show you than over-claim and gate you. Running and proven right now, on a real v1.0.0 deploy: multi-provider Chat with persistent, searchable memory; the no-code Flows builder with signed run traces; multi-provider model routing (Ollama / Vertex / OpenAI / Azure / Bedrock), with a different provider per capability if you want — Bedrock for chat, Vertex for images; semantic three-layer tool selection over the bundled MCP servers; in-chat agentic tool execution — the model calls a real MCP tool mid-conversation, behind a human-in-the-loop approval gate on anything that mutates, and the result comes back in-thread; a tamper-evident audit trail; per-tool credential isolation; and one-command install to Docker Compose or Kubernetes. Every claim here has a receipt — a screenshot or a live recording — on the Check it out page. Not a sign-up form.
The lock-in era is ending — our open-source core gets you there sooner
Here's the part the rented players don't want said out loud: the moat is evaporating. Capable open models already run on a laptop, and the gap to the frontier shrinks every month. Soon any model you'd want will run locally — and when the model is a commodity you host, vendor lock-in becomes a thing of the past. Our open-source core is built for that world now, so you can start moving toward it today instead of waiting for permission: swap providers per task, run frontier APIs when they earn it and local models when they don't, and never rewrite your workflows to do it.
And it's not just about cost or control — it's about capacity. Our open-source core and our platform dramatically scale what a team can do. They let a novice, non-technical operator drive their cloud more safely and more capably than almost any human could alone — by combining the strengths of multiple models, grounding every answer in your real systems, and acting through real MCP servers that securely run your AWS, Azure, GCP, and Kubernetes. Multiple model "brains," real grounding, and least-privilege tools that act as you — pointed at the stack you already run. For a team under a budget, a compliance regime, and a deadline, that combination is a no-brainer.
Why this is the whole point
The reason so much of this category is a meter is that the business model is the dependency. Keep it in their cloud, keep the model theirs, keep your data and your identities on their servers, and you keep paying. "Agentic" becomes a subscription.
Our open-source core removes the dependency. You decide which model, where it runs, who it runs as, and how much you spend — because you host it yourself, on your own identity, on your own infrastructure. That's not a smaller version of the rented tools. For the people who actually have to run this in production — under a budget, under a compliance regime, under their own control — it's the only version that matters. And it makes "request early access" look like what it is.
It's free to run. It runs as you. It's running today. Install it in one line and see for yourself:
curl -sSL https://install.openagentics.io | bash