Kivali

---

Most AI assistants keep one memory: a short list of facts about you, shared across every chat session. Kivali gives you colleagues instead of sessions: durable agent personas, each with a role, its own memory and its own workspace, organised into a team that reports to you.

A gardener that knows your garden. A researcher that remembers every paper it has read for you. An operations lead that knows which supplier is always late. Each one gets better at its job the longer it works with you, and no change to the team happens without your say.

What could your team run?

A team is for anything with more moving parts than one person can keep in their head. Every team starts with a Chief of Staff who helps you hire the rest.

A small business or startup

Hire the team you can't afford yet.

• Marketing lead · knows your customers and how you sound
• Bookkeeper · knows how you categorise every expense
• Operations lead · tracks every supplier and order
• Researcher · keeps an eye on competitors

"Marketing lead, work with the bookkeeper on a spring promotion we can afford, and bring me the plan to approve."

A home or home office

Hand over the admin that never ends.

• House manager · knows the warranties, the paint colours and when the boiler was last serviced
• Family planner · meals, trips and the school term
• Back office · invoices, receipts and tax paperwork for your freelance work

"Back office, our energy bill doubled. Pull the last two years of bills and work out with the house manager what changed."

A side project or community

Give the thing you love a team of its own.

• Researcher · remembers every source it has read for you
• Builder · has its own computer to write and run code
• Writer · drafts the posts, newsletters and grant applications

"Researcher, find out what our members want from a club app, then work with the builder on a first version I can try on Saturday."

Ever thought I wish I had someone for that? Now you do.

Why Kivali

Durable, not disposable

Every agent has a lasting identity: a role, a curated memory of what it knows, habits it has learned about how to work, and a digest of every past conversation. When a chat grows long, you start a new one: the agent first folds what it learned into its memory and habits, so it keeps the knowledge and sheds the clutter.

You stay in charge

You are the head of the team. Agents bring you approvals and updates in one inbox. New hires, role changes and handbook edits are proposals you approve. Messages between agents wait in a queue that you release, by hand or on a delay you choose, so the team moves at the pace you set, and so does what it costs.

Yours, on your machine

Kivali runs on your computer or your own cluster. Every piece of state is a plain file you can read, back up and restore. Agents work in sandboxed shells with an allowlist for the internet, and you bring your own Claude: a subscription, an Anthropic Console account, Amazon Bedrock, Google Vertex AI or Microsoft Foundry.

What your team can do

• Learn your world. Drop in plans, specs, notes and decks (PDF, Office documents, Markdown, images) when you create a team. They set the context for every agent you bring on.
• Grow as an organisation. The Chief of Staff manages hiring, role changes and reorganisations. You approve each change, and it takes effect a moment later.
• Take on real work. Goals break down into assignments with owners, parts and dependencies. Agents wake when something they own changes, ask for what they need as assignments of their own, and close the loop when they are done.
• Work in parallel. An agent can fan a job out to focused subagents that run side by side and report back.
• Use real tools. Each agent has its own Linux shell to run code, process data and build things, with internet access limited to the sites you allow.
• Build shared knowledge. Agents publish their work: reports, designs, requirements and decisions. Published files form a knowledge graph that every agent can query, so the team stays consistent about what has been decided.
• Remember. Agents keep curated memory, learned habits and searchable digests of past conversations. You can read and edit what each one remembers.
• Learn new skills. Install skills (reusable instructions and scripts) for the whole team, and turn each one on or off.
• Show you the bill. Every model call is metered, per agent and per day, so you always know what the team costs.

How it works

flowchart LR
    you(("You"))
    inbox["Inbox<br/>approvals · updates"]
    cos["Chief of Staff"]
    team["Your agents<br/>role · memory · shell"]
    work[("Assignments")]
    kg[("Knowledge graph")]

    you -- "approve · reply · release" --> inbox
    inbox <--> cos
    cos -- "proposes hires<br/>and changes" --> inbox
    cos <--> team
    team <--> work
    team -- "publish" --> kg
    team -- "need a decision" --> inbox

You talk to any agent directly, and agents talk to each other through messages you release. Work is tracked as assignments. Anything that needs a yes or no lands in your inbox. Read How Kivali works for the full picture.

Quick start

Kivali Desktop (macOS and Windows)

Kivali Desktop runs on macOS 13 or later on Apple silicon, and on 64-bit Windows 10 or 11 Pro, Enterprise or Education with Hyper-V turned on.

1. Install it.
   • macOS: download Kivali, open the .dmg and drag Kivali to Applications.
   • Windows: turn on Hyper-V if it is not on already (Turn Windows features on or off, Hyper-V, restart), then download Kivali and run the installer. It asks for administrator rights once, to install the service that runs your teams' virtual machines.
2. Open Kivali and create a team. Kivali Desktop sets up a private virtual machine for the team and starts it.
3. Sign in with Google, then connect Claude: a Claude subscription, an Anthropic Console account, Amazon Bedrock, Google Vertex AI or Microsoft Foundry.
4. Follow the setup wizard: upload a few documents about your work and hire your Chief of Staff.

The getting started guide walks through each step.

Your own Kubernetes cluster

Download the chart and the images for your architecture from the latest release, then, on k3s:

zstd -dc kivali-images-linux-amd64.tar.zst | sudo k3s ctr images import -

kubectl create namespace kivali
kubectl -n kivali create secret generic kivali-secrets \
  [email protected] \
  --from-literal=session-key="$(openssl rand -hex 32)"

helm install kivali kivali-X.Y.Z.tgz -n kivali \
  --set externalURL=https://kivali.example.com

See Self-hosting for other clusters, sign-in, connecting Claude, networking and upgrades.

Try it without installing

From a clone, make run starts a demo team at http://127.0.0.1:8080 with a scripted stand-in for the model, so you can look around without signing in to Claude. You need Go and Node; see the developer guide.

Documentation

Get started	Getting started · How Kivali works · FAQ
Use it	Using Kivali · Memory · Assignments · Knowledge graph · Kivali Desktop
Run it	Self-hosting · Configuration · Sign-in · Backup and restore · Security and privacy · Troubleshooting
Build it	Developer guide · ·

Contributing

Issues and pull requests are welcome. Start with CONTRIBUTING.md and the developer guide. To report a security problem, follow SECURITY.md.

License

Kivali is licensed under the Apache License 2.0; see also NOTICE. Each release ships the notices of the third-party software it includes in THIRD_PARTY_LICENSES.txt, and where to get the source of its GPL and LGPL components in SOURCES.md.