What an artefact is, and why your agents make so many
An artefact is the finished thing an agent hands back: the pricing page, the board deck, the research memo, the chart of last quarter's numbers. Not the chat that led to it, and not the edits in your repository. The thing itself, the one you would send to someone.
The thing at the end of the task
Ask an agent to do a job and two things come out. There is the work in progress: the conversation, the false starts, the scratch files it wrote along the way. And there is the result, a single file you can open, read and pass on. That result is the artefact.
In myartefacts, each artefact is one file with a few things attached to it: a short human title, a one-sentence summary, the request that produced it, the agent that made it, the project it belongs to and a version number. It also gets a stable slug, so it keeps the same link however many times it is revised.
- A pageA landing page, a pricing table, a small app with its own scripts.
- A deckSlides built as one HTML file, or a .pptx to download.
- A documentA Markdown brief, a Word report, a plain text note.
- DataA CSV export, a TSV table, a JSON file.
- A pictureAn SVG diagram, a PNG chart, a generated image.
- A PDFAn invoice, a one-pager, a paper to read later.
The kind is worked out from the file's extension: HTML, Markdown, SVG, Image, PDF, Data, Doc, Deck or File. An agent can say more when the extension cannot, for example that an HTML file is really a Deck or an App. The extension also decides how the file opens, and every format myartefacts shows is covered one by one in the next article.
Scratch files, logs and changes to code inside a repository are not artefacts. Those belong to your disk and to git. The instructions myartefacts gives agents say exactly this: publish when you finish a standalone artefact, not for every file you touch.
Why there are so many
Agents are cheap to ask and quick to answer, so you ask more. You want three takes on the landing page, so you get three files. You ask for the same chart with a different axis. A research task ends with a memo, a table of sources and a short summary for your team. One morning with two or three agents can leave a dozen artefacts behind, and the people who work this way run agents every day.
Then they scatter. Each agent keeps its output in its own place: a chat history here, a session folder there, a file in Downloads called final-v3.html. Sharing means emailing an HTML file around, or sending a link on the vendor's domain that shows the tool rather than you. Finding something two weeks later means remembering which agent made it, on which day, in which window.
Most of the pile is drafts, and that is fine. Drafts are how you get to the good one. The real job is to spot the few worth keeping, and to put your hand on them again later without digging.
One library for all of them
myartefacts is a private home for everything your agents make, whichever agents you use. Each one connects with its own token, and its name becomes the source shown on everything it publishes, so work from Claude Code and Codex sits side by side and you can filter by either.
- Every day is a sheet. New artefacts land on today's contact sheet as numbered frames with live previews, newest first.
- Every artefact has its own link. Publishing again under the same slug adds a new version at the same URL, so the links you handed out keep working. See Update an artefact without breaking its link.
- Agents file their own work. Projects nest up to three deep and each one says what it is for, so an agent reads the list and picks the right place. See Projects: how agents file their work.
- Private until you say so. Nothing is public by default. When you do share, the link can expire, can let visitors save the file, and counts who opens it. See Share an artefact.
- Search covers what you asked. Agents send the request that produced each artefact, so search looks through titles, slugs, summaries, prompts, projects, agents and kinds. Try a word you remember typing.
Keep the few, let the rest go quiet
The library is built for a fast pass through a pile, the way a photographer goes over a contact sheet with a grease pencil. Everything works from the keyboard:
- ␣ opens a frame in the loupe, full size, with its summary and the prompt that made it.
- S stars a keeper.
- F archives a frame. It leaves the sheet but stays searchable, its link keeps working, and F again brings it back.
- C copies its link.
- / jumps to search.
Starred, Shared and Archived each have their own view beside the sheets, so the keepers are always one click away and nothing you set aside is lost.
Getting your first artefact in
The quickest way is to let an agent do it. Connect Claude Code in one command, or set up Codex, Cursor or any other MCP client. Once an agent is connected, it publishes without being reminded each time: the server tells it to publish anything standalone it makes, file it in a project, and reply with the title and a link that opens it in your library.
When the file is already on disk, the agent sends it with a single curl command rather than pasting it into a tool call, as Publish a file that is already on disk explains. And you can always drop a file onto the library yourself: any file or folder, up to 20 MB each.