Prompt it.
Ship it.
Iris is your always-on AI orchestrator — managing agents, running workflows, and executing tasks while you focus on what matters. Any model. Any tool. Fully yours.
iris cloud — Coming Soon
Skip the VM setup, Terraform configs, and infra babysitting. Deploy a fully managed, hardware-isolated agent operator in minutes — not infrastructure sprints.
Running agents in production
is still mostly duct tape.
Agents die when your session ends.
No persistent runtime means every agent restart is manual. Context is lost. State lives in terminal history.
No one knows what agents are allowed to do.
You can't audit what your agent accessed, what it changed, or what rules it was following. No chain of custody.
Every model switch is a migration.
Lock-in to one provider breaks your entire agent infrastructure when pricing changes, rate limits hit, or a better model ships.
Iris is the answer to all three.
From zero to a running agent fleet in minutes.
Prepare the environment.
Set up an Ubuntu 22.04 VM, install dependencies, create a Slack app, and gather your API keys for the LLM provider, Slack, GitHub, and optionally Azure.
Clone and bootstrap.
Clone the irisflow repository, run bootstrap.sh, and let it automatically install Docker, Node.js, Terraform, the GitHub CLI, and configure the runtime service.
Configure runtime and secrets.
Choose your secrets backend, a local .env file or Azure Key Vault, configure the model and provider settings, and verify that iris.service starts successfully.
Enable Firecracker isolation.
Optional advanced setup — requires a KVM-capable VM (Azure Ddsv5 series). Run bootstrap.sh --firecracker to install Firecracker, build the VM root filesystem, provision sandbox VMs via Terraform, and connect Iris to a static VM or dynamic pool.
Run and manage Iris.
Message @iris in Slack, use skills and sub-agents, manage sandbox VMs, swap models and providers, and maintain the system with Terraform, hot-reloaded skills, and GitHub as the single source of truth.
A persistent engine.
Not a chatbot.
Iris sits between you, the LLMs, and your infrastructure. It runs as a system service, reading guidelines and coordinating a fleet of sub-agents to execute your workflows.
Host mode
Runs directly on your host VM. No KVM or nested virtualization needed. Best for private, trusted instances.
Firecracker sandbox
Spawns isolated microVMs per Slack channel. Fresh state every session. Ideal for untrusted code execution.
No virtualization overhead
A runtime with a heartbeat
Iris is managed by systemd, auto-restarts on failure, and executes actions directly in isolated sandboxes.
CONSTITUTION.md
Hard operator rules injected read-only before every prompt. Version-controlled. Cannot be overridden at runtime.
You decide what your AI can do.
Tell Iris what you want her to do, and what it should never do. It'll follow your rules, every time.
Whether you want a co-pilot that drafts PRs for your review, or a fully autonomous agent that deploys fixes at 3 AM, Iris scales with your trust.
Your AI, your rules. No surprises.
Run specialized agents for any workflow.
Newsletter agent
Iris monitors your sources, drafts your weekly newsletter, and waits for your approval before sending. Set it and forget it.
If you can prompt it,
Iris can run it.
Agents for sales outreach, customer support triage, research synthesis, calendar management, contract review, any repeatable workflow you can describe, Iris can own.
Iris comes ready. Gaps, it fills herself.
The built-in skills cover secrets, storage, Terraform, GitHub, Azure, agent spawning, and more. Every skill hot-reloads without a restart and when Iris encounters something it can't do, it extends herself and it's live on the next turn.
Engineering
Infrastructure
Communication
Productivity
What engineers are saying.
“Iris is the control plane we didn't know we needed. Went from a tangle of scripts to a single orchestrator running in Slack. Night and day.”
“The Firecracker isolation is what sold us. Every agent gets a fresh microVM. Nothing bleeds between sessions.”
“We replaced three orchestration tools with Iris. One VM, one service, one place to look when something goes wrong.”
“Hot-reloading skills without a restart sounds small until you've shipped a fix to a live agent in 10 seconds.”
“The GitHub-as-source-of-truth model is exactly right. The VM is disposable. The repo is everything.”
“bootstrap.sh on a blank VM and Iris is running in under five minutes. That first @iris mention in Slack still feels like magic.”
“Iris is the control plane we didn't know we needed. Went from a tangle of scripts to a single orchestrator running in Slack. Night and day.”
“The Firecracker isolation is what sold us. Every agent gets a fresh microVM. Nothing bleeds between sessions.”
“We replaced three orchestration tools with Iris. One VM, one service, one place to look when something goes wrong.”
“Hot-reloading skills without a restart sounds small until you've shipped a fix to a live agent in 10 seconds.”
“The GitHub-as-source-of-truth model is exactly right. The VM is disposable. The repo is everything.”
“bootstrap.sh on a blank VM and Iris is running in under five minutes. That first @iris mention in Slack still feels like magic.”
“Swapping from GPT-4o to Claude mid-project with a single env var change. No redeploy, no code change. The provider abstraction is genuinely well done.”
“The constitution model is underrated. Hard operator rules injected before every prompt, version-controlled, impossible to override. That's the right way to govern agents.”
“Iris writes her own skills when it hits a gap. Watched her self-extend in production for the first time and just stared at the screen.”
“MicroVMs booting in 125ms. We're running dynamic sandboxes per Slack channel and the overhead is negligible.”
“Self-hostable, MIT licensed, no vendor lock-in. For the first time I actually trust the AI infra we're running on.”
“The VM is disposable and a full rebuild takes one command. Most underappreciated thing about Iris. Sleep better at night.”
“Swapping from GPT-4o to Claude mid-project with a single env var change. No redeploy, no code change. The provider abstraction is genuinely well done.”
“The constitution model is underrated. Hard operator rules injected before every prompt, version-controlled, impossible to override. That's the right way to govern agents.”
“Iris writes her own skills when it hits a gap. Watched her self-extend in production for the first time and just stared at the screen.”
“MicroVMs booting in 125ms. We're running dynamic sandboxes per Slack channel and the overhead is negligible.”
“Self-hostable, MIT licensed, no vendor lock-in. For the first time I actually trust the AI infra we're running on.”
“The VM is disposable and a full rebuild takes one command. Most underappreciated thing about Iris. Sleep better at night.”
Stop building agents.
Start running them.
Iris is MIT licensed, lightweight, and fully open-source. We built the agent orchestration engine we always wanted to use - no bloat, no lock-in, just a runtime that works.