uploads.sh MCPサーバー

The missing upload command for coding agents.

Capture screenshots, recordings, and other artifacts as you work: test reports, logs, JSON, PDFs, zips. When the pull request opens, uploads.sh puts them in one tidy comment that updates automatically on each revision. Hosted uploads.sh is free to start. Connect your own bucket — Cloudflare R2 or any S3-compatible provider — so storage in that bucket is unmetered, or self-host the open-source service.

---

Artifacts ready when the pull request opens

uploads hosts the artifacts coding agents produce at stable public URLs they can use in pull requests and issues: screenshots and screen recordings, but also test reports, logs, JSON, CSV, Markdown, PDFs, and zip archives (the full list is under Plans & limits). On a branch, uploads put stages each file as soon as it is ready. When the pull request opens, uploads.sh promotes the staged files into one managed comment.

GitHub's own attachments work from a browser and, since GitHub CLI 2.99 (September 2026), from gh … --attach, but only once a pull request or issue exists, and the files stay behind GitHub's own hosting, where they cannot be embedded or fetched anywhere else. uploads.sh gives agents a stable public URL from the same terminal where they build and test the change, while the branch is still in progress; GitHub has no public API (yet) for that. The hosted MCP server does the same for agents that cannot run a command at all, taking files as bytes or a URL.

Keys are hash-free, so re-uploading the same filename overwrites in place and the URL never changes — every embed of it updates at once. Workspaces keep tenants (and their budgets and key policies) apart.

This repo is the source of the canonical deployment at uploads.sh: the API worker, auth worker, MCP server, the hosted read-gate Worker, the Astro web app, and the @buildinternet/uploads CLI (published to npm from packages/uploads).

What it looks like

One comment per PR, rewritten in place on every sync. Files tagged --state before and --state after pair into a side-by-side table; other images and video land below it, and non-media files (reports, logs, archives) list in a file table with their type and size.

The real comment on #436.

Pairing is by --meta path=… when several pairs share a comment (one before and one after per path), and falls back to filenames that differ only by a before/after token — hero-before.webp with hero-after.webp.

Everything you attach also lands in your workspace, grouped by where it came from — pages by path, projects by repo or app — and browsable from one place.

The screenshots view groups uploads by project and path.

Open any file for a share page: copy-ready embeds (Markdown, HTML, and more), the raw URL, and a delete button.

Each file's share page — copy-ready embeds, details, and delete.

Quick start

Install the CLI and sign in once:

npm install --global @buildinternet/uploads
uploads login

Upload a file and tag the page it shows:

uploads put ./settings.png --meta path=/settings

On a branch, put stages the file automatically. Open the pull request however you normally would. The GitHub App promotes the staged files into one managed attachments comment.

More ways to upload

Use the same commands for before-and-after evidence, an open pull request, a browser capture, or an annotated image:

# Pair two states from the same page in the pull request comment.
uploads put ./before.png --meta path=/settings --state before
uploads put ./after.png --meta path=/settings --state after

# See what this branch will attach when the pull request opens.
uploads staged

# Attach files directly when a pull request or issue is already open.
uploads attach ./before.png ./after.png

# Capture, annotate, and upload a page in one command.
uploads screenshot http://localhost:4321/settings --via local --annotate ./callouts.json

attach detects the repository and current PR through gh, uploads all files, and creates or updates that same one comment. On an issue it also keeps a Markdown issue artifact (and a gallery link, when there is one) unless the workspace or the repo turns that off. Without the GitHub App, run uploads attach --promote after opening the pull request to promote files that you staged earlier. All commands run under npx @buildinternet/uploads … without a global install.

Sign in with GitHub or a magic link, then create your own workspace or accept an invite into one — see enrollment. Hosted files are public URLs — private-repo attachments get non-guessable links (how that works), but anyone holding a URL can view the file. A file that looks like it shows a credential, ID, or card number is held for your review before it goes public, but don't rely on that: do not upload secrets or sensitive UI.

Require screenshots in CI

The GitHub App adds an uploads.sh check to pull requests that change UI files, next to your tests and builds. No workflow or Actions step is needed. Out of the box it only reports, so nobody is blocked. Turn on Require screenshots on UI changes in the workspace's Settings → GitHub, or set it for one repo in .uploads.yml, and the check fails until screenshots arrive:

comment:
  check:
    require: [before-after, desktop, mobile] # or just [uploads]

The PR comment shows a checklist that ticks itself as files arrive. Images dragged into the PR on GitHub count too, so contributors without uploads.sh are never stuck, and a skip-uploads label skips a PR that needs no screenshots. Add uploads.sh as a required status check in branch protection to make it block merges. Setup, which files count as UI, and skip labels are in Require screenshots on pull requests.

Connect your agent

The hosted MCP server runs at https://agents.uploads.sh/mcp and is listed in the MCP Registry as sh.uploads/mcp. Local stdio is uploads mcp on the same npm package.

# Claude Code
claude mcp add --transport http uploads https://agents.uploads.sh/mcp

# Codex
codex mcp add uploads --url https://agents.uploads.sh/mcp

# OpenCode
opencode mcp add uploads --url https://agents.uploads.sh/mcp

uploads install adds the agent skills and the MCP server, so future sessions can capture each visual milestone without being asked. The skills also install standalone into any agent runtime:

npx skills add buildinternet/uploads

That installs three skills: github-screenshots (visuals → PRs/issues), uploads-cli (full CLI reference), and annotate-screenshots (callouts and redaction on a capture).

On Claude Code or Codex, the plugin bundles the skills, the MCP server, and a pre-PR screenshot reminder. In Claude Code it also shows staged screenshots above the prompt and attaches them when the PR opens. On claude.ai and in Cowork, add uploads from the Claude directory. The same folder (plugins/claude/uploads) is the common agent-plugin layout, and it is submitted to the xAI plugin marketplace as uploads.

Full CLI usage, including annotations, managed comments, public galleries, and change feeds (repo-wide or one pull request), lives in docs/cli.md.

uploads feed create prints a live page of the screenshots on a repo:

uploads feed create --repo buildinternet/uploads.

REST routes are in docs/api.md.

What's in this repo

Path	What
apps/	The deployables: the REST API worker (api.uploads.sh), the optional public viewer, the auth worker, the remote MCP server, the Slack app (link previews and held-upload messages), the hosted read-gate Worker, and the Astro site at uploads.sh
packages/	Shared code — most notably @buildinternet/uploads (the CLI, published to npm), @uploads/storage (the files-sdk adapter factory all storage goes through), and @uploads/plugin (Claude/Codex plugin version, not published)
skills/	The three agent skills that ship to users
hooks/, plugins/, .mcp.json	Agent-runtime wiring: the shared pre-PR screenshot hook and the Claude / Codex plugin manifests
server.json	MCP Registry listing (sh.uploads/mcp): stdio uploads mcp plus the hosted remote

Each worker and the web app deploy separately. All storage access goes through createStorage() in packages/storage — adding a provider is one new case plus peer deps, no API changes.

Hosted file bytes stay on the shared uploads-default bucket. store.uploads.sh, storage.uploads.sh, and embed.uploads.sh are served through uploads-read-gate (apps/read-gate). A held image gets a "Held for review" card. Any other safety hold is a 404. Every other request passes through to R2. It does not gate private files, and it does not run for a bring-your-own bucket. Operator notes are in docs/ops.md.

Docs

Product docs — install, the staged loop, the GitHub App, limits — live at https://uploads.sh/docs. The docs in this repo are the companion: CLI and API reference, contributor setup, and operator material, all mapped from docs/README.md.

How to set up, test, and open a pull request: CONTRIBUTING.md. Where the project is headed: VISION.md. Agent working conventions live in AGENTS.md, and agents that land on this repo should start at llms.txt. The product site serves https://uploads.sh/llms.txt and https://uploads.sh/llms-full.txt.

Local development

Prerequisites: Node ≥24 and pnpm ≥11 (corepack enable). No Cloudflare account needed for the core local loop — wrangler dev simulates R2, KV, and D1 on disk:

pnpm bootstrap        # one-command setup: tooling, deps, env vars, local D1, default workspace
pnpm dev              # API on :8787 (local R2 + KV + D1)

bootstrap is idempotent, and pnpm doctor diagnoses a setup without changing it. The rest of the loop — the authenticated dev stack, the check and test gates, and how to open a pull request — is in CONTRIBUTING.md.

License

Apache 2.0.