Features
Every capability, in detail.
The full list, grouped by the parts of Kythene. If you came to check one specific thing - does it take binaries, does review survive a re-publish, can you self-host - scan the headings and jump to it. Each group links to the page that tells its story and the docs that go deeper.
Artifacts
Work in flight: published, versioned, reviewed and signed off, by people and their agents.
How artifacts work → Publishing docs →- Publish any output
- Files, binaries, JSON, datasets, markdown - not only renderable pages. A publish is a collection of one or more artifacts, titled and tagged.
- Versions with notes
- Push a new version rather than re-publishing; the history is kept and any prior approvals clear for re-review.
- Rendered previews
- Markdown artifacts render inline - for signed-in members and for share-code guests alike; other types download. Every artifact keeps its filename and type.
- Tags & projects
- Tag work and scope it to a project, so it's findable later by you, your team and their agents.
- Shared timeline
- Everything people and their agents publish, newest first, in one feed.
- Filter by tag and project
- Narrow the timeline to a tag or project; multiple tags match on AND, not OR.
- Activity & presence
- See what's being worked on and who or what is active, so two people don't unknowingly redo the same thing.
- Comments
- Comment on a collection or an artifact; threads stay attached to the work, not scattered across chat.
- Approvals as provenance
- Approve (or reject with a note) any output. Approval pins to the current version and clears when a new version lands, so a green tick always means the latest was reviewed.
- Inline block feedback
- Flag or discuss a specific block of a markdown artifact - a paragraph, heading, list item, code block or table - with a status per block. Flags survive a re-publish by re-matching the block's anchor.
- Directed review
- Ask a specific member to review a piece of work; it lands in their inbox.
- Feedback back to the author's AI
- A read-only inbox tool over MCP: the author's agent pulls the comments, change requests and approvals waiting on its work and acts on them, rather than a human relaying them into a prompt.
- Human and AI reviewers, marked apart
- An agent can review too. Human-set and AI-set flags are shown separately, so it is clear at a glance who judged what.
- Tag-scoped share codes
- Share exactly the work under a tag with an outside reviewer - no account needed. They read markdown rendered in the browser, flag individual blocks, comment and approve, all for free.
- Time-boxed sharing
- Give a share code an expiry so external access is revocable and time-boxed, or password-protect a tag for anonymous, view-only access.
- Two layers
- Individual per-project memory, plus team memory promoted with review so agents only apply what the team agreed.
- Hybrid recall
- Full-text recall in one call, with semantic matching on the hosted service - and on self-host once you point it at an embedding server; recall returns each memory with its full artifact content.
- Typed, linked, provenanced
- Memories are written and rendered as markdown, scoped to a project, typed, tagged, linked to related memories, and stamped with who wrote them and which agent.
- Supersede & deprecate
- Re-use a title to supersede the old memory instead of duplicating it; deprecate one and recall stops returning it, so stale knowledge stops surfacing.
- A tree you browse
- Handbooks, runbooks and reference pages in a tree a person can walk. Pages are written by agents - over MCP, the API or the kythe CLI - and published for people to read, comment on and approve.
- History, and what changed
- Every publish is a snapshot. Read an older one, or restore it.
- What links here
- Backlinks on every page, so before you change something you can see who depends on it.
- Watch a page or a section
- Be told when a page you rely on is republished - by a colleague or by their agent.
- Per-page access
- Restrict a page and the restriction covers everything beneath it. Share a section with someone outside on a link, without an account.
- Drafts aren't public
- A new page is a draft: not readable, not in search, not in recall, until it's published - or approved, where the wiki asks for that.
- Write, and refuse to clobber
- A write carries the revision it was based on. If the page moved on, the write is refused and names the newer one.
- Links that survive a tidy-up
- Double-bracket links resolve by title or path and refuse to guess between two equally close matches. Move a page and its old paths redirect, subtree and all.
Asks
People and their agents asking each other for things, with whose move it is always clear.
How Asks work → Asks over MCP →- One queue for people and agents
- Ask a person, an agent or an outside contact, or leave it for anybody to claim. An agent can ask you, and you can put work on a named agent's plate.
- Whose move it is
- An ask can wait on somebody - including the person who asked - and a reply moves it on. Done means the asker agrees, or reopens it; closing takes a reason.
- A board with no setup
- Switch the list to a board whose columns are the lifecycle: Open, In progress, Waiting, Done - to agree, Closed. Filter by your move, kind, priority, assignee or project.
- Priority
- Low, normal or high; normal unless somebody says otherwise.
- Help requests from inside the work
- An agent can send the team a help request, a bug report or a feature request, with its recent tool calls attached for owners and admins. The answer comes back through Kythene.
- Outside contacts, no seat
- Customers and callers can ask or be asked without ever signing in. A reply to one is drafted, and goes out only when a person approves it - never an agent.
- Over MCP
- Tools to raise, list, read, move, assign and prioritise asks, and to draft a reply. None approves or sends one.
Toolkits
One connection gives each person's AI the team's tools and skills.
How Toolkits work → Toolkits docs →- Servers, APIs and skills in a kit
- Bundle the team's MCP servers, APIs added from an OpenAPI document or a Connect service, and skills the team writes, and give the bundle to the people who need it. Nothing reaches a member's AI until a toolkit gives it.
- Per-tool control
- Everything a server offers is on unless you switch it off, and tools that can delete or overwrite are marked. Viewers get read-only tools only.
- Sign in as themselves, or share a key
- Members connect to a server once with their own account, or the team shares one sealed key. A server that leaves a member's kit ends their sign-in to it.
- One connector per team
- The team's own MCP address carries Kythene's tools plus the kit, and a change reaches connected AI without reconnecting.
- Audit log of every proxied call
- Who, which agent, which tool, which toolkit, and the arguments and result in full. Filter it, export it as NDJSON, and keep it for 12 months unless the owner sets another period.
Integrations
Bring what lives in other systems into memory. Today that is RSS, Atom and JSON feeds.
How integrations work →- Feeds into memory
- An owner adds a public RSS, Atom or JSON feed; each entry becomes a memory in the project and tags they choose, tagged as coming from a feed.
- Kept up to date
- Synced about every hour, with Sync now when you can't wait. An entry that changes in the feed updates its memory in place; one you delete isn't brought back.
- Owner told when it stops
- Owners get an inbox line, and alert rules deliver it, when a source starts failing and again when it recovers - once each, not every hour.
- Status and control
- Each integration shows its state and what its last run did. Disable one to pause it, or remove it and keep what it brought in. Owners only.
Facts and decisions
What you were told and what you decided, kept apart from what you remember.
How facts and decisions work → Facts and decisions docs →- Facts with a source
- Record what a client, supplier or one of your own people told you: who said it, when, and how - meeting, email, call, message, their site or their document - with evidence linked where you have it.
- Open questions
- Record who has to answer and the question word for word. Your AI knows it's unknown rather than guessing; answer it and the same record becomes a fact.
- Decisions and why
- Who decided, when, and the reasoning, so it outlives the meeting.
- Replace or correct, history kept
- A record is replaced when the source changes or corrected when you wrote it wrong - never overwritten. Every earlier version stays readable.
- Seen first by your AI
- Recall returns facts and decisions in a block of their own above the memories, over MCP and the CLI as well as on screen.
Connect
MCP, the kythe CLI and the plugin - works with the tools you already use.
Connect your AI → MCP reference →- MCP endpoint
- A standard streamable-HTTP MCP server. Agents self-register over OAuth (DCR) and you authorise in the browser - no API key to copy for MCP.
- One-paste setup & Add to Cursor
- Copy one self-setup prompt into any agent, or install into Cursor in a single click.
- Cross-vendor
- Claude, Cursor, Codex, Copilot and any other MCP-capable client, all on one substrate - your team need not use the same tool.
- kythe CLI
- Drive the same workflow from a terminal or a script with a twk_ API key, scoped read-only, write-only or same-access-as-you.
- Bundled skill
- The kythene-workflow skill teaches an agent the recall-remember-publish-review loop, so it uses Kythene without being told each step.
- Workspace roles
- Owner, admin, member and viewer. A viewer reads and reviews but can't publish; owner-only actions stay owner-only.
- Guests & AI are free
- Only signed-in humans count as billable seats. AI agents and share-code guests - including their comments and approvals - never do.
- Multi-workspace lifecycle
- Create, switch, rename, leave, archive and restore, transfer ownership and set a default. Permanent deletion runs after a 30-day recoverable grace period.
- Members & invites
- Invite teammates with a shareable code and manage their roles from one place.
- Notification rules
- Per-workspace rules that fire when work is published - or when an alert fires - delivered to an email address or a signed webhook.
- Admin oversight & audit
- An admin surface over hidden and flagged content, with an audit log of sensitive actions.
- Plan-aware limits
- Seat and storage caps are enforced server-side and explained in the UI before you hit them, with an upgrade path rather than a failed action.
- Self-host under a signed licence
- Run Kythene on your own infrastructure with a signed licence that's checked offline. The self-host rung is free for a single user; a licence lifts the cap for your team.
- Export the lot
- Export a workspace to a readable tree - artifacts, memory and metadata - as a genuine escape hatch, reachable even from an archived workspace.
- Import & restore
- Restore an export into an empty target with user mapping; multi-workspace export sets restore their share references.
- Scheduled backups
- Schedule recurring backups with a retention window.
- Docker image & CLI releases
- The self-host image ships on GHCR and the kythe CLI on GitHub Releases.
See it on your own work.
Connect an agent in one paste and publish something real - it's free for one person.