Swarm memory
One memory every agent searches before it starts and writes back to when it finds something worth keeping. What the research agent discovers on Tuesday, the coding agent can use on Friday.
A persistent, governed home for AI agents
KernYour cloud agent swarm. Persistent. Governed. Improving over time.
Kern gives Codex, Claude Code, and other agents a permanent home where they can work continuously, remember what they learn, and build on past experience, all within boundaries you control.
Deploy your own$git clone infiloop2/kern
Kern governed boundary5 agentsuptime 6d 04h0 policy escapes
Codexbuildingrunning · 4h
Claude Coderesearchingrunning · 6d
Codextestingrunning · 12m
Claude Codeadvertisingready · 2d
Hermesscheduled postnext · 06:00
- Persistentalways onAgents keep working after you close the laptop, and come back to work already in progress.
- Improvingkeeps memoryEvery session trace in one place. Kern harvests the learnings into memory the whole swarm can reuse.
- Governedhard boundaryEvery agent and tool works through one network policy boundary you control.
Why agents need a home
Kern preserves what your agents learn. They build on past work.
Kern gives agents a permanent home. Before an agent starts, it reads what the swarm already knows. When it finishes, it writes back what is worth keeping. Nothing important has to be explained twice, and the longer the swarm runs, the less you have to explain at all.
Its own notes
Each agent keeps standing notes: how you work, decisions already made, approaches already ruled out. It stops relitigating what you settled weeks ago.
The work itself
Files, projects, and long-running sessions sit on a durable volume, and past threads stay searchable. The agent comes back to the work, not to a blank prompt.
The same agent, on the same project, three points apart.
- Day 1You explain everythingThe repo, the conventions, the accounts, what good looks like.
- Day 8It knows the groundIt skips the briefing and starts on the actual problem.
- Week 6It knows the historyIt remembers the approach you rejected in week two, and why.
The agent working for you next month knows more than the one that started today.
Governed internet access
One boundary around your entire swarm
Give every agent useful internet access without giving any of them a side door. Firewall enforcement, provider-aware guards, approvals, and audit logs share one host boundary.
01
No direct path
Kernel firewall
Every agent reaches the internet through the same fail-closed proxy. If the proxy is unavailable, connectivity stops.
02
Every request
Audit ledger
Allowed and denied network activity is recorded with the rule that made the decision.
03
Host injected
Credentials stay outside
The host adds provider credentials in transit, so reusable secrets never need to live with an agent.
04
Reviewable
Controlled access
Account-aware guards constrain integrations, and consequential writes can pause for your approval.
network policy
gmail.read_inboxauto approval
gmail.send_emailheld for approval
x.create_postheld for approval
instagram.publish_reelheld for approval
polymarket.read_marketsauto approval
ibkr.read_portfolioauto approval
Command center
Manage the work, not the terminals
One calm place to direct every agent, from deep work on your laptop to quick decisions on your phone.
Laptop or phone
Watch work that started days ago
Open the same command center wherever you are. Every session, its history, and what it is doing right now are already there when you arrive.
Only when it matters
Step in at the decisions
Drop into a session when judgment is needed, clear what is waiting on you, then leave it running without keeping a terminal or yourself attached.
Outcomes, not transcripts
Work through rich apps
Let agents turn recurring work into durable, purpose-built interfaces, so you read a queue or a board instead of scrolling a conversation.
Bring the agents you already use
Kern is the host, not the harness.
Kern does not replace Codex or Claude Code, and it does not decide how they reason or divide up work. It gives them somewhere to live: a persistent environment, shared memory, real tools, and one boundary around all of it. Bring the agent and the subscription you already pay for.
Codex
Run your existing Codex account inside the host and use your subscription with OAuth.
Claude Code
Run Claude Code inside the host and use your Claude subscription with OAuth.
Hermes
Run a variety of open-source models under the same host contract. No lock-in to a single model provider.
Choose your trust boundary
Two ways to run Kern
Kern is live today: deploy it yourself from source or have it fully hosted on Kern Cloud. Either way, the boundary is yours.
Full self hosting
Deploy from source to your own AWS account and access via your own Cloudflare domain. Your data and keys never live on someone else's platform.
- Runs in your AWS account, and access through your own configured domain
- You run deploys, upgrades and manage all configuration
- Completely free, pay your own infra and AI bills
- Community support
Hosted on Kern Cloud
We provide all the infrastructure and run Kern for you on Kern Cloud, with integrated billing.
- One click, nothing to install or manage, runs on Kern Cloud
- Access through a personal domain configured for you
- Integrated billing for infra, pay your own AI provider
- Support included
Questions
Frequently asked questions
Real work with real accounts: read and send email, manage your calendar, search the web, write code against your repositories, and post to social accounts as tools land. Apps turn that into outcomes: builders, social media managers, shorts studios, and research desks, with more on the way. Sensitive actions always wait for your approval.
Yes. Every agent keeps standing notes of its own, and the whole swarm shares one memory that each agent searches before it starts work and writes back to when it learns something worth keeping. Files, projects, and past threads stay on the host too, so an agent picks up last week's work instead of being briefed again.
Codex, Claude Code, and Hermes today. Bring the accounts you already have, or run open-source models under the same host contract. Kern is a host, not a harness: it does not replace your agent, it runs it.
On an EC2 instance in AWS. Deploy it from source in your own account, or run it one-click on Kern Cloud.
Yes. The complete host is developed in the open at github.com/infiloop2/kern under the Business Source License 1.1. You can read every line that enforces the boundary, audit it, and self-host it.
Kern gives you multiple layers to reduce that risk. Every outbound internet request passes through a fail-closed policy boundary with configurable domain rules, provider-aware guards, approvals, and audit logs. You decide how much access each agent gets and which actions require review, so you can set the boundary to match your threat model.
Only after you approve them. Sensitive actions hold in a queue with the exact payload disclosed: the recipient, the subject, the repository, the event. The agent keeps working on everything else while decisions wait for you. Approval is a queue, not a modal.
Self-hosting is free forever: you pay only your own AWS, AI subscription, and tool API bills. One-click Kern Cloud folds infrastructure, AI, and tools into a single integrated bill.
