ArtifactBefore
Useful output with nowhere durable to live.
calculator-final-v3-really-final.html sits in Downloads. The source is buried in a chat, ownership is unclear, and the next edit replaces the last copy.
Versioned artifact vault · public alpha
Keep the artifact. Keep the source. Keep every version.
ArtifactFlow is a self-hosted, versioned artifact vault for tools and documents created with AI. Keep their authoritative source, find them later, preview them, run generated HTML away from the authenticated app, and preserve every revision.
Built first for executable HTMLKeep generated single-file HTML tools searchable and versioned, then run their untrusted code on a separate origin away from the app that holds your login.
Current alpha: single-file HTML, Markdown, and Mermaid · AGPL-3.0-or-later · built for internal teams.
Preserves the output.
Not the conversation.
Source firstAuthoritative, portable content
Version nativeHistory, diffs, restore, revert
People + MCPOne permission and audit model
Runnable HTMLSeparate origin, no app cookies
The gap after generation
The missing artifact layer between AI chat and production.
Chat is temporary. A new repository, deployment, and release process are too much for every calculator, one-pager, runbook, diagram, or prototype. ArtifactFlow gives useful output a durable place in between.Built for technical teams that regularly create small internal tools, operational documents, diagrams, and prototypes with AI.
Preserves the output, not the conversation.
Not a vector database, chat archive, static site generator, or agent memory system. ArtifactFlow keeps the deliberate artifact that should survive after the chat ends.
After
A managed team asset without another software project.
The same calculator has a stable identity, authoritative source, retained versions, ownership, permissions, search, safe preview, and a path for the next human or agent update.
The artifact lifecycle
From generated output to a reusable team asset.
People can bring work in through the browser. Approved AI clients can preserve it through MCP. Both routes reach the same versions, permissions, search, and audit trail.
Keep the authoritative source.
Save a self-contained HTML tool or portable Markdown document with its title, owner, workspace, tags, and category.
Search, preview, and run the result.
Find artifacts by metadata or content. Read documents in place and let interactive HTML run inside its isolated preview boundary.
Preserve the next version too.
A person or agent updates the source without erasing history. Inspect source diffs, restore an earlier version, and preserve which provider/model produced the content when that identity is known.
Terminology: An artifact is the managed record. Each version retains its authoritative source or original; previews and runtimes are derived ways to use it.
One vault, two interfaces
One vault for people and agents.
ArtifactFlow is not an AI generator. It is the system of record that remains after the model moves on.
For approved agents
Retrieve the source. Preserve the revision.
MCP clients can search, read, create, update, and revert within explicit token and workspace scopes, while declaring exact version-level provider/model provenance when known.
For people
Search, inspect, preview, run, and govern.
The web interface is the human control and consumption layer, including editing, workspace organization, permissions, ownership, and account security.
01
search
query: "Q3 capacity"
1 result · Q3 Capacity Plan
02
read
page_uid: 01ky7atfyh…0hef0f
source + current_version_uid
03
update
base_version_uid: 01KY7ATFYM…53H9NN
new immutable version preserved
Static transcript
Captured from real MCP calls against an isolated test database. Identifiers are shortened for display; the full session is on the MCP page.
The product today
Search the vault. Preview the artifact. Keep working.
The screenshots show the product as it exists today, not concepts, mockups, or a future roadmap.
Executable when useful
Run generated HTML without turning it into a deployment.
Interactive calculators, dashboards, diagrams, and runbooks remain usable, while untrusted code executes away from the authenticated application that holds the team's login and data.
Authenticated app
The place that knows the team.
Search, source, versions, workspaces, permissions, and audit stay on the application origin. Artifact HTML never executes there.
Isolated artifact runtime
The place that assumes the code is hostile.
HTML runs in an opaque sandbox on a separate cookieless origin. Ordinary network requests are blocked, with documented self-navigation and browser-dependent WebRTC residuals.
Honest boundary:
Scanning and the injected guard are bypassable defense in depth. Browser-enforced origin separation, the iframe sandbox, and response-header CSP carry the security boundary.
Understand the boundaries
Follow each concern to its source.
Detailed mechanics, operational requirements, product direction, and engineering provenance are documented separately.
Artifact workflowStable identity, immutable versions, draft status, metadata changes, and format-specific payloads.
MCP for agentsScoped retrieval and versioned writes through the same application rules.
Security modelWhy useful HTML can run without executing where users authenticate.
Self-hostingLocal evaluation, production roles, origins, storage, and boot checks.
Engineering harnessHow ArtifactFlow became the real security-sensitive workload for testing AI engineering guardrails.
Product roadmapPDF and Word document focus, proof requirements, and explicit non-promises.
Public alpha · July 2026
Built, tested, and still young.
Pin a revision. Expect breaking changes. Read the threat model before trusting it with work that matters.
- Executable artifacts are currently limited to single-file HTML. Network access is intentionally constrained by the runtime security model.
- No public sharing, marketplace, approval workflow, or built-in AI generation.
- Source diffs exist; a visual diff UI does not yet.
- Internally adversarially reviewed and broadly tested, but not independently audited by a third party.
After the chat ends
Run ArtifactFlow locally.
Evaluate the artifact workflow and security boundaries against your own use case.