THE MESH
Reference
Back
Cloudflare Blog · blog.cloudflare.com ↗ Beta

Artifacts: Git for Agents

A Git-compatible versioned filesystem built agent-first, running on Cloudflare infrastructure.

Artifacts provides a distributed, versioned filesystem for agents — not as an adaptation of existing git hosting, but as a ground-up implementation designed for the scale and patterns of agentic workflows. Tens of millions of repositories, created programmatically, with fork/diff/merge semantics available to any piece of logic that needs them.

Architecture

The system runs on Cloudflare Durable Objects with a custom Git server implementation written in Zig, compiled to WebAssembly (~100KB). Storage is layered:

Crucially, it speaks native Git protocol (v1 and v2) — shallow clones, incremental fetches, git-notes for agent metadata. Any standard Git client can authenticate via HTTPS with a generated token and treat it like any remote.

Access surfaces

Native Git
Any git client works. Clone, fetch, push via HTTPS with token auth. Full protocol support including shallow clones and incremental fetches.
TypeScript SDK
Direct programmatic access from Workers. env.AGENT_REPOS.create(name) creates a repo. Then treat it like any git remote.
REST + Workers API
For serverless environments without a git client. Full CRUD on repos, files, and refs.
ArtifactFS
Filesystem driver with blobless clones and lazy file hydration. Reduces large repo startup from ~2 minutes to 10–15 seconds.

What it's designed for

Relevance to the Mesh
Artifacts maps directly onto several Mesh requirements. The canonical layer (P3) could be Markdown files in an Artifacts repo — versioned, Cloudflare-native, agent-first, with no need for a separate GitHub sync (P4). Fork semantics are the exact primitive needed for the protected-page review workflow in P8: draft = fork, review = diff, approve = merge. Section-level addressing and incremental fetches (P10, P12) map to file-level git operations.
Status: Beta — to explore further
Artifacts is in public beta as of May 2026. Worth evaluating as the canonical storage layer for the Mesh before committing to a D1/KV-based design. Key questions: how does section-level addressing map to file granularity (one file per section? one file per page?), and what does the auth model look like for multi-user access at different scope levels (org/team/individual).

Pricing (beta)