Products and open source
Iris is an orchestration runtime for AI agents.
A persistent harness that manages agents, runs durable workflows, and executes tasks across any model and any tool. The Iris Stack adds the governance and quality layers that make that autonomy trustworthy. Open source under Apache 2.0.
Join the Iris Cloud waitlist(opens in a new tab)iris-core on GitHub(opens in a new tab)
What it's for
Iris exists because of the things that actually break AI systems in production, not because of what looks good in a demo. Agents that die between sessions and lose everything they were holding. No audit trail for what an agent actually did on your behalf. Behaviour welded to one provider's API, so the system's economics and availability are somebody else's roadmap.
It answers those with a persistent runtime, durable workflows that survive restarts, human-in-the-loop approvals as a state the system can wait in rather than a meeting that happens afterwards, and safe retries and recovery.
How it runs
Iris runs on your own VM or in isolated Firecracker microVM sandboxes. Autonomy is a dial rather than a switch — Supervised, Guided, Informed, Autonomous — so how much the system is allowed to do on its own can expand as the evidence justifies it, and contract when it doesn't. It's reachable over Slack and Telegram.
Why it's open
People adopt infrastructure they can audit. Code that others can read, run, and criticise gets held to a higher standard than anything built behind a wall — and it lets the people building the same future check our work instead of taking our word for it.
A managed Iris Cloud offering is coming.
Products and open source
The Iris Stack
Production AI systems need more than an agent that can act. They need a runtime that persists, a clear moment for human judgment, and proof that changes have not made the system worse. We build each layer independently, so teams can adopt one or use them together.
- 01
Runtime
Iris Core
The persistent, self-hosted runtime for agents that work across Slack, Telegram, and the web; use any model, any tool, and isolated sub-agents when the work calls for it.
- — Durable workflows
- — Agent orchestration
- — Model and tool independence
Open source · Apache 2.0
- 02
Governance
RelayKit
The human-in-the-loop relay layer inside Iris. It routes agent decisions to the right person for approval, missing context, review, or takeover — with a full audit trail behind every handoff.
- — Agent-to-human relay
- — Approval states
- — Audit trail
Private beta · Repository details on request
- 03
Quality
Pupil
Continuous quality engineering for AI agents. Pupil runs repeatable evaluations, measures real operating signals, and catches regressions before a changed prompt, model, or workflow reaches production.
- — Scenario evaluation
- — Regression gates
- — Quality metrics
Open source
Coming soon
Iris Cloud
The managed, cloud-hosted way to run Iris Core. We are building it for teams that want the runtime without having to operate the infrastructure themselves.
Portfolio
Also building
Running our own products is how the engineering practices get tested before they reach a client. These are at varying stages.
GoTixi(opens in a new tab)
Launched, scaling
An AI-first travel company.
CELRYS
Ideation
Testing, building, deploying, and monitoring software in an AI-driven development world — designed for agents as much as for humans.
Omee(opens in a new tab)
Brewing
A creative studio where kids and teens turn ideas into digital worlds, without needing to code.