Getting started

Fifteen minutes from nothing to your agent recalling context you gave it earlier. The step that matters is connecting an agent. The rest happens in conversation.

This assumes the hosted app at https://kythene.com. Self-hosting? Swap in your own URL.

1. Sign in

Sign in with GitHub or Google at kythene.com. Your first sign-in creates a personal workspace - the container for your work, members and billing. It starts as just you.

• Team: create or join a shared team workspace and invite people from Settings -> Members. You can belong to several.
• Memory stays in the workspace you're working in. A finding in your personal workspace is yours until you promote it into the team workspace, where an owner approves it before the team's agents recall it.
• Promotion moves a memory between two workspaces you belong to, so the team shape needs both.
• Projects organise work inside a workspace. A project is just a tag, so they're free: make one per repo, client or initiative. See Core concepts.

2. Connect your AI agent

This turns Kythene on. A connected agent can publish, recall, remember and review without you leaving the session.

The fast way - one paste. Give this to any agent that speaks MCP (Claude Code, Claude desktop, Cursor, Codex, Copilot) and it connects itself:

You are my AI agent. Connect yourself to Kythene - my team's shared memory and
output layer - by fetching https://kythene.com/setup-prompt.md and doing exactly
what it says. Then tell me it worked.

No config file, no API key. The agent registers over OAuth and you approve it in a browser tab.

Cursor users get a one-click Add to Cursor button on the Kythene home page.

Wiring it up by hand. Your MCP endpoint is on the Connect your AI page in the app:

https://kythene.com/mcp/kythene

• --scope user makes the connection available in every project and session. Without it Claude Code scopes it to the current directory, and a new session (below) finds no Kythene tools.
• The first Kythene tool call opens a browser to sign you in and authorise the connection. No API key is needed for MCP.
• Your team needn't agree on one agent. Everyone connects what they already use, and all of them share the workspace.

Claude Code users: take the plugin. One install sets up the connection and teaches your agent the workflow: recall at session start, remember what it learns, publish, pick up feedback. In Claude Code:

/plugin marketplace add kythene/plugin
/plugin install kythene@kythene

Self-hosting? Point it at your own instance instead:

/plugin marketplace add https://your-instance.example.com/.claude-plugin/marketplace.json
/plugin install kythene@kythene

The archive install needs Claude Code v2.1.224 or newer. On an older version, download kythene-plugin.zip from your instance and run /plugin marketplace add ./kythene from the unzipped folder.

There's more on what it teaches in the skill reference.

3. Publish something

Ask your agent, in whatever words you like:

"Publish that benchmark you just ran to the api project, call it Nightly benchmark."

It creates a collection (one publish, one or more artifacts) tagged api and gives you a link. In the app you'll see the file rendered, its version number, and who and which agent produced it.

Try a markdown document too. It renders inline, so someone who never opens a terminal can read it.

4. Remember something, then recall it

Your future sessions lean on this. Tell your agent something worth keeping:

"Remember for the infra project: the migration has to run before the worker boots, or it deadlocks on the advisory lock. Call it Deploy gotcha."

Then start a new session, on any machine, and ask:

"Recall what we know about deploys on infra."

The gotcha comes back with its title, project, author and agent, before your agent repeats the mistake. That's the whole product in two prompts.

Recall returns published work as well as memory by default, so "what do we know about X" pulls back artifacts too.

5. Set up your workspace guide

Every workspace has a short operating manual that connecting agents read automatically: house style, your tag taxonomy, what belongs in memory. New workspaces start from a sensible default.

Ten minutes in Settings -> Workspace guide pays off the first time an agent writes a memory the way you would. See The workspace guide.

6. Optional: the CLI

For scripts, cron jobs and CI. The kythe CLI uses an API key, not the browser login.

Download the archive for your platform from github.com/kythene/cli/releases, extract the binary and put it on your PATH:

tar xzf kythe_*_$(uname -s | tr A-Z a-z)_*.tar.gz kythe
sudo mv kythe /usr/local/bin/
kythe --help

Create a key on Connect your AI (or Settings). It's shown once. Then:

export KYTHE_URL=https://kythene.com
export KYTHE_TOKEN=twk_...

Full command list in the CLI reference.

Where next

• The model, once: Core concepts.
• Pages your team and your customers read: Wikis.
• The way you actually work: pick a walkthrough.
• Running it yourself: Self-managed hosting.