An early public preview, told as a story
Hermes Artifacts
Agents save their work privately. A guest link is a separate step.
Agents make things. A report with a chart, a diagram, a small HTML tool that does one job. Hermes Artifacts is where those things go after the chat scrolls on: one private library for all my agents, in the Hermes dashboard, and a separate, deliberate way to show one of them to someone else.
Early public preview 0.1.0: install from source, there's no package or release yet. Source on GitHub · Where it stands
Chapter one
Chat is a bad place to keep a treasure
An agent's best output is often a finished thing: a one-page report with a chart, an SVG diagram, a small HTML calculator. In a chat it turns into an attachment or a wall of code. To read it properly I download it. To find it next week I scroll. To show it to someone I forward a file or paste it somewhere public, and then I can't take it back.
Artifacts gives that output a place of its own, inside the dashboard I already sign in to, and keeps sharing as a separate decision. The rest of this book follows one report from the chat to a guest and back.
Chapter two
The relic goes into the library
When I ask for something finished, the agent calls artifact_publish with a file path, or with inline text and a kind. The artifact lands in the Artifacts library in my Hermes dashboard, and the reply says, in so many words, that no link exists. Hermes keeps plugin tools behind its tool search, so the model looks them up first.
An artifact never changes. Publishing again makes a new artifact, the way a relic, once forged, is never reforged: a second one is made beside it. Formats are self-contained HTML, Markdown, SVG, PNG and WebP, up to 10 MiB each.
Every agent and profile pointed at the same service sees the same library. It lists everything newest first, filterable by title, kind and the profile that published it. Each artifact has its own private address, like /artifacts/latency-review, taken from its title. That address names the artifact and grants nothing: it only works behind the dashboard's sign-in. Open one and it gets the workspace: inline, the full browser window, or fullscreen. If I'd rather not open Hermes, an optional standalone console shows the same library behind its own login.
Chapter four
A guest at the window
The guest needs no sign-in. They get no library and no other files. The artifact fills the window; a small menu in the corner shows its title, when the link expires and a Download link. The guest site lists nothing, so there is nothing to wander into.
A link that expired, was revoked, never existed or is mistyped gets the same plain 404, so it gives nothing away. Even the 404 doesn't say which of those it was.
Chapter five
The sword cuts the string
From the dashboard, or with artifact_revoke_share and the share's id, I take a link back. That link stops at once. Other links to the same artifact keep working, and my copy stays in the library.
No sword reaches into someone else's sketchbook, though. Anything the guest already downloaded stays downloaded. Revoking stops the link, not the copy.
Chapter six
Every relic lives under glass
Agents write HTML, and HTML can carry a prompt injection or a page copied off the web. So every artifact is treated as hostile. Nothing an agent publishes ever runs in the dashboard's own page, and the guest site never sees dashboard credentials.
- A sealed frame. Scripts may run, but in an opaque origin: no cookies, no storage, no reach into the page around it.
- No network. The content policy blocks fetches, external images, form posts and frames. Self-contained means self-contained.
- No wandering off. If an artifact tries to navigate its frame elsewhere, the viewer puts it back and says so.
- A separate guest site. It answers only reads of one artifact; the shipped proxy config strips cookies and
Authorizationbefore it.
Chapter seven
Six charms on the belt
An agent gets six tools, all in the artifacts toolset. Without the service URL and credential configured, each one answers not_configured and the dashboard page shows a setup hint instead of an empty shelf.
| Tool | Does | Parameters |
|---|---|---|
artifact_publish | Saves a file or inline text privately. Creates no link. | title, path or content + kind |
artifact_list | The library, newest first, metadata only. | query, kind, profile, limit ≤ 100, cursor |
artifact_get | One artifact, its links with status, a text preview. | artifact_id |
artifact_download | Writes the original bytes to a local file. | artifact_id, path, overwrite |
artifact_share | One guest link for one artifact. The URL is returned once. | artifact_id, expires_in_hours (168) |
artifact_revoke_share | Revokes one link. Others keep working. | share_id |
Chapter eight
What the magic won't do
Every storybook has rules about its magic. These are the ones that matter before you trust it with anything.
- It can't recall a download. Revoking stops the link, not the copy a guest already saved.
- A link is a key, not an invitation. Whoever holds it can open the artifact until it expires, including whoever it was forwarded to.
- The sandbox isn't an air gap. The content policy doesn't cover WebRTC, and browsers differ on DNS prefetch, so a hostile artifact could still signal out. Don't share sensitive content as an artifact.
- “Share only when asked” is wording, not a lock. The tool description tells the model so. What enforces it is that sharing is its own call, and every link shows up in the dashboard, where I can revoke it.
- Five formats. HTML, Markdown, SVG, PNG and WebP, up to 10 MiB each. No PDF, no video, no multi-file sites; HTML can't load anything from outside itself.
- One library, one operator. Every agent and profile configured for the service shares one library. A profile is where an artifact came from and a filter, not a separate account. The console has one operator login, with no per-user accounts.
- An agent can publish what it can read. A prompt-injected agent could publish a secret file it can reach. Sharing it would still take a separate call that shows up in the dashboard and can be revoked.
- Chromium only, so far. The sandbox and content policy are tested in Chromium. Firefox and Safari are expected to fail closed, but nobody has tested them yet.
- Old Safari gets nothing. Browsers that don't send
Sec-Fetch-Dest(Safari before 2023) are sent to the shell and can't show guest content. It fails closed.
Chapter nine
Where the story stands
An early public preview, 0.1.0. The source is public and works; it hasn't been production-tested beyond one private pilot, and interfaces and the database schema may still change.
Done
- The service, the standalone console and the Hermes plugin, with their test suites.
- A release gate on 2026-10-09: browser tests against an isolated Hermes 0.21.5 dashboard in headless Chromium, a container build and restart, and a real agent publishing through the plugin.
- One private pilot behind a reverse proxy, with the guest site on its own bridge.
- A launch review by two independent read-only reviewers, with the fixes in the security notes.
Not yet
- No package, container image or release. Run it from source.
- Tested against Hermes 0.21.x only, and in Chromium only.
- An older database is refused, not migrated.
- No third-party security audit.
To try it, the quickstart runs the service on your own machine with one command. Then install the plugin into Hermes. Read the security model first.
Appendix
The technical part
Controls, defaults and the two runtimes, for readers who skipped the pictures. Open what you need.
Security controls, one row per threat
Every artifact payload is treated as untrusted. Agents write HTML, SVG and Markdown, and that content can carry prompt injection or copied web pages.
| Threat | Control |
|---|---|
| Artifact code running in the dashboard | Artifacts never enter the Hermes page. They run in sandbox="allow-scripts" frames with an opaque origin: no allow-same-origin, popups, forms, top navigation or downloads. |
| Network exfiltration | The content CSP is default-src 'none' with connect-src 'none', form-action 'none', base-uri 'none' and frame-src 'none'. Inline script and style only; media only from data: and blob:. |
| Markdown and SVG | Markdown renders with raw HTML off and unsafe link schemes dropped. SVG renders inside <img>, so its scripts don’t run. |
| A frame navigating itself away | A trusted wrapper frame owns the artifact frame’s navigation, restores the artifact and shows a notice. |
| Link leakage | 128-bit random tokens (22 URL-safe characters), stored only as SHA-256. Referrer-Policy: no-referrer, Cache-Control: no-store, noindex. Access logs write /<share> instead of the token. |
| Links that fail open | One resolver checks format, existence, revocation and expiry for the shell, the content and the download, on GET and HEAD. Every failure is the same 404. |
| Anonymous management | The management API needs a machine credential, checked before the body is read and again by the router. The guest-only bridge never forwards it. |
| Credentials in the dashboard | The machine credential stays on the server. The Artifacts tab asks Hermes for one of a fixed set of API requests, made with the dashboard’s own session. |
| Anonymous standalone console | Off unless configured. Then the page, its API, previews and downloads all need the operator’s bcrypt-checked login, and a misconfigured login stops the service from starting. |
| Resource exhaustion | 10 MiB per artifact, list pages of at most 100, a per-client rate limit (IPv6 grouped by /64), container memory and process limits. |
Read-only reviews by AI reviewer agents during development, and a launch review by two independent read-only reviewers on 2026-10-09, found no critical or high-severity issues. They are reviews, not a third-party audit. The highest-rated findings (a preview frame that could navigate itself away, and the content URL opened as a top-level page) are fixed. Some were accepted as residual: CSP doesn’t cover WebRTC or DNS prefetch, “share only when asked” is the tool’s wording, not a policy, an agent’s file tools can publish any file it can read, and the service needs a proxy in front that times out slow clients. The full list is in the project’s security model.
Defaults and limits
| Setting | Default | Notes |
|---|---|---|
| Size of one artifact | 10 MiB | Checked by the tool before upload and again by the service. |
| Guest link lifetime | 7 days | expires_in_hours on artifact_share, up to 8,760 hours. The service caps it at 365 days by default. |
| Library page size | 20 | limit on artifact_list, 1 to 100, with a cursor for the next page. The HTTP API defaults to 30. |
| Public requests | 120 per minute | Per client; IPv6 clients are grouped by /64. |
| Formats | HTML, Markdown, SVG, PNG, WebP | HTML must be self-contained: inline CSS and JS, no external URLs. |
Three parts
The service owns artifacts, links, rendering policy and the guest viewer: one Python 3.14 FastAPI process with SQLite. It runs on its own, so guest links keep working while Hermes restarts. Its management API is reachable only with a machine credential that never leaves the server. Agents outside Hermes can use that API directly.
The standalone console is optional: the same library page at
/console/, behind one operator login, off unless configured.
The Hermes plugin runs inside Hermes 0.21.x. It adds the six agent tools, a packaged operating skill, the Artifacts tab in the dashboard and a thin adapter behind the dashboard’s own sign-in. It holds the service credential on the server side and stores no data itself.
The guest viewer is a read-only site: it renders one artifact per link and
lists nothing. It can’t publish, delete or manage links, and the shipped
proxy config strips cookies and Authorization headers on the way in.