webfetch.
The license-first image layer for AI agents and humans. One API, CLI, and MCP that federates 24 licensed sources, falls through to a human-like browser, and always ships attribution.
$ npm i -g @webfetch/cli
Not stock photos. Real works, fetched live, stamped with provenance.
Every image below was fetched from a public source with a verified license. Hover to see attribution and a link back to the origin. This is what your agent sees.














































One config line. Your agent now fetches licensed images.
claude mcp add webfetch -- npx -y @webfetch/mcp
- search_imagesfederated search across 24 providers
- fetch_imagedownload + write XMP sidecar
- get_licenseresolve license tag for any URL
- list_providerscapability + rate introspection
- search_audiocover art + album metadata
- attribution_blockship-ready HTML / Markdown caption
We don't ship URLs. We ship attribution-ready output.
A short history of the night sky

Long before telescopes, before satellites, before we knew that the smear of light across the summer sky was our own galaxy seen edge-on, people painted what they saw...
Attribution lives next to the image, not in a spreadsheet. The sidecar becomes your figcaption.

Why we go.
— Carl Sagan
Export attribution-safe hero slides from a single MCP call. The credit line is generated, not copy-pasted.

The Hope Diamond
E-commerce and marketing pages get a visible license pedigree — no legal-scramble before launch.
Shipping an image used to take an afternoon.
Six failure modes, fixed at the protocol layer.
Candidates are sorted by license tag (CC0 > PUBLIC_DOMAIN > CC_BY > CC_BY_SA > EDITORIAL), then metadata confidence. UNKNOWN is rejected by default.
Wikimedia, Openverse, Unsplash, Pexels, Pixabay, NASA, Smithsonian, Met Museum, LOC, iTunes, MusicBrainz CAA, Spotify — one interface.
When public APIs miss, an opt-in managed browser pulls from Google Images and Pinterest. Every result ships an attribution sidecar.
One config line installs into Claude Code, Cursor, Cline, Continue, Roo Code, and Codex. Your agent gets a stable fetch tool it can reason about.
Every download writes an XMP sidecar with license, creator, source, rights, and SHA-256. Audit-ready by default, no extra work.
Local runs phone home to nothing. Cloud telemetry is opt-in and pseudonymous. No third-party trackers, ever.
One core. One cache. Every surface.
Every surface shares @webfetch/core. The cloud router adds metering, pooled keys, and a managed browser you never have to operate.
24 federated providers.
One ranked stream.
Licensed sources rank first. Editorial-licensed sources next. UNKNOWN providers are strictly opt-in and always emit an attribution sidecar.
Every result carries a real license, sourced from a real provider, with a real link back to the original.
webfetch vs alternatives
Full comparison →| Tool | Commercial safe? | Aggregates sources? | Attribution? | MCP native? | Self-host option? | Browser fallback? | Usage-based pricing? |
|---|---|---|---|---|---|---|---|
| webfetch | yes | yes | yes | yes | yes | yes | yes |
| Google Images (retired API) | no | no | no | no | no | no | no |
| Unsplash API | yes | no | partial | no | no | no | no |
| Bing Image Search | partial | no | no | no | no | no | yes |
| Serper / SerpAPI | no | partial | no | no | no | yes | yes |
Start free. Pay for the parts you can't build.
OSS unlimited on your machine. Managed browser, pooled keys, and audit logs in the cloud.