Screenshots and recordings in GitHub PRs and issues
Why this exists
GitHub's native image hosting (github.com/user-attachments/…) is reachable
from a browser and, since GitHub CLI 2.99 (September 2026), from
gh issue|pr create|edit|comment --attach <file>. There is still no public REST
endpoint (the CLI uses an undocumented upload host), the upload needs push access, and the hosted file is private to
GitHub. Any other image URL in a PR/issue body written with gh … --body-file
must already point at something publicly hosted. The uploads CLI and the
hosted MCP at https://agents.uploads.sh/mcp both host the file on uploads.sh
and return a stable public URL plus ready-to-paste markdown.
When gh --attach is enough
Use plain gh pr comment --attach ./shot.png (or create/edit) when ALL
of these hold:
gh --versionis 2.99 or newer and you have push access to the repo.- The PR or issue already exists, or you are creating it in the same command.
- The file is an image or video under 10 MB (video up to 100 MB on paid plans).
gh --attachtakes media only. - Nobody needs the link outside GitHub.
user-attachmentsURLs do not render in Slack, docs, or other tools, and other agents cannot fetch them.
Use uploads.sh when any of these are true instead:
- The artifact is not media: a Lighthouse or test report, a log, JSON, a PDF, or
a zip.
gh --attachdoes not take those;uploads putdoes, and the managed comment links them. - The PR does not exist yet. Stage on the branch with
uploads put; the attachments comment appears when the PR opens. - The shot will be re-taken. Same filename, same URL, and the managed comment updates in place instead of piling up.
- The capture needs metadata (
--meta), later retrieval (uploads find), or the Screenshots view across projects. - The URL must work outside GitHub.
- You are on the hosted MCP with no shell, lack push access, or the repo is on GitHub Enterprise Server.
Do not mix the two on one PR: pick one so the comment history stays readable.
Media attached with gh is still indexed: the GitHub App imports attachments
from PR and issue text into the workspace. To pull an existing
user-attachments image out to a public URL by hand, use uploads ingest.
Which surface
Pick one transport and stay on it. This skill is the workflow. The uploads-cli skill owns flags and MCP tool contracts.
| You have | Use |
|---|---|
| No shell (ChatGPT, or any host without a checkout) | Hosted MCP put with (contentBase64 + filename, or contentUrl) + repo + (pr or branch). Filename is optional with contentUrl when the URL path has a leaf. Embed the returned markdown / embedUrl. Never imply you can run uploads attach ./shot.png. |
A checkout and the uploads binary |
The CLI examples below. Git can fill repo / branch. |
A localhost page or a selector annotate |
CLI only (uploads screenshot --via local). Remote render cannot reach your machine. |
| Neither MCP nor the CLI | Stop and say so. Do not treat npm install -g as the ChatGPT path. |
On the hosted MCP there is no attach tool and no git defaults. Stage with
put + branch + repo. Once the PR exists, promote with repo + pr +
branch, or put with pr + repo (optional branch also promotes). The
managed comment is bot-only on that server.
Step 1 — Capture the visual
Skip this step if the visual is already in context (a ChatGPT attachment, a
file the host already holds). Go straight to hosted MCP put.
Prefer uploads screenshot <url|file.html> — it captures and hosts in
one step (drives a local Chrome, or falls back to a server-side render), so you
skip a separate host call. It takes --viewport WxH@Nx, --wait, --selector,
--full-page, and --out <file> (to also save the PNG).
Capturing your own dev server? No manual DOM surgery needed:
- Known framework dev toolbars (Astro/Next/Nuxt/Vite) are hidden automatically
— opt out with
--no-hide-dev-tools. --reduced-motionsettles animations.--hide <selector>hides any other overlay (repeatable).--eval <js>/--init-script <file>(local backend) are the escape hatch to dismiss a banner or freeze a specific animation.
Capturing a clicked/selected state of a React/Next (or other hydrating) app?
A synthetic el.click() in --eval fires before the framework hydrates — the
element is in the SSR HTML but no handler is attached yet, so it silently does
nothing. Gate on the app's own "interactive" signal with --wait-for <js>
(local backend), which polls that expression until truthy before --eval runs:
uploads screenshot http://localhost:3000 --via local \
--wait-for 'document.querySelector("[data-hydrated]")' \
--eval 'document.querySelector(".tab-settings").click()' --out settings.png
uploads screenshot http://localhost:4321 --viewport 1520x960@1x --out home.png --reduced-motion
uploads screenshot https://uploads.sh --selector main --dark
--out also drops a <file>.uploads.json sidecar next to the PNG — that's a
working file for a later put/attach to pick metadata back up from, not
something to commit, so .gitignore it (*.uploads.json) or delete it once
you're done attaching.
Only reach for your harness's browser tools / Playwright / an existing file when
uploads screenshot can't reach the target (e.g. a flow that needs auth or
interaction first). GIFs and video: capture with any tool and upload as-is — the
optimizer only rewrites still images (PNG/JPEG → WebP).
Step 2 — Host and embed
Two tiers, pick by whether a PR already exists:
- Simple — a PR/issue already exists.
uploads put shot.png --pr 123(oruploads attach shot.png, which infers the PR from the current branch) — one call, stable per-PR key, embed URLs back immediately, and the managed comment collects that PR's media as a side effect. - Advanced — stage pre-PR, before there's anything to target.
uploads put shot.pngon a branch (see below) — no PR/issue needed yet; promotion and the comment happen automatically once the PR opens, but only for a repo already bound to the workspace (see the caveat below). Reach for the simple tier once the PR exists unless you're deliberately building up a staged set across a longer branch.
Default loop: stage as you go, from the first visual milestone. Don't wait for a PR to exist — the moment you have something worth capturing, mid-task on a branch, attach it right then.
Two commands stage automatically, no extra flag needed:
- A bare
uploads put, whenever you're inside a git repo on a non-default branch with no--pr/--issue/--key/--ref/--prefix. It stages under the same branch-keyed pathattach --branchwould produce. uploads screenshotwith no--pr/--issue/--branchtarget. Capturing straight from a URL carries every derived fact — path, url, env, viewport, plus--state— through to the PR once it opens.
Reach for attach --branch explicitly only when you want its extras: uploading
several files at once with shared flags, or triggering promotion/comment sync as
a side effect.
Staging keys embed the branch name. Renaming it is followed automatically the
next time you run any uploads staging or promote command on that branch — if
the PR opens before that happens, recover the stranded files with
uploads attach --pr <n> --from-branch <old-branch-name> (see uploads-cli).
uploads screenshot http://localhost:4321/settings --out step1-before.png --state before
uploads screenshot http://localhost:4321/settings --out step2-after.png --state after
# or, capturing an existing local file instead of a live URL:
uploads put ./step1-before.png --meta path=/settings --state before
# or, explicitly, e.g. to upload several at once:
uploads attach ./step1-before.png ./step2-after.png --branch --state after
This uploads under stable, branch-keyed paths (no PR/issue target needed, no
comment yet — there's nothing to comment on until a PR exists). Keep doing
this at each meaningful visual milestone as you work; don't batch everything
into one call at the end. On the default branch (or outside a git repo, or
with --no-git), put/screenshot fall back to their ordinary dated
layout — that's the opt-out, along with any explicit --key/--ref/
--prefix/--destination.
Staging only auto-promotes into a bound repo — don't promise it blind. Auto-promotion at PR-open time (webhook or CLI-triggered, below) needs the repo already bound to a workspace. Binding happens two ways:
- Implicitly, from any earlier successful attach/comment/promote call against that repo.
- Explicitly, via
uploads github link.
A repo that's never been bound and only ever staged with --branch sees no
error and no comment when the PR opens — a silent no-op. So if you can't
confirm the repo is bound (uploads github link --status), don't tell the user
the screenshot will "just show up." The fallback that works regardless of
binding history: once the PR exists, run uploads attach --promote (or any
targeted uploads attach against that PR) to promote and post explicitly.
Pass --state before/--state after and --meta path=/route as a habit —
both, every time. They're the two highest-value queryable tags, and the two a
tool can't recover later:
--statebefore/after is the whole point of most PR screenshots, and nothing can infer it from the image.pathis just as easy to forget, except onuploads screenshot, which derives it from the captured URL. Aput/attachof an existing file has no URL to derive it from.
Both cost one flag now and make uploads find state=after or uploads find path=/settings work months later, when the filenames mean nothing to anyone.
For a same-URL before/after pair, uploads screenshot is the straightforward
path. Its object name derives from the captured URL, and --state folds into
that name — localhost-docs-mcp.webp becomes -before.webp/-after.webp. So
capturing the same URL twice with different states lands two distinct objects
instead of one overwriting the other. Pass an explicit --key when you need a
specific object name. A put of an existing file has no URL to derive a stem
from, so keep its before/after distinct with separate filenames or --key:
uploads screenshot https://app.example/settings --pr 123 --state before
uploads screenshot https://app.example/settings --pr 123 --state after
uploads put ./after.png --pr 123 --meta path=/settings --state after
(--state also takes empty, error, and loading.) When an image lands with
no path meta, attach/put --pr/put --issue print tip: add --meta path=/route so this shot is findable by page on stderr (and a JSON hint
field) — don't ignore it. Viewport is derived for you on screenshot. See the
uploads-cli skill for the full canonical vocabulary.
The PR comment assembles itself — you don't drive that step. Once the PR
opens (whether via gh pr create or the GitHub UI), every branch-staged file
gets promoted into that PR's attachments and the managed comment is created
automatically. Files you never mention in the description still show up there.
The comment is identified by a hidden HTML comment at the top:
<!-- uploads.sh:attachments ws=<workspace> -->. Bot posts are from
uploads-sh[bot]; the local-gh fallback uses the same marker.
- With the uploads-sh GitHub App installed on the repo, a webhook does this the moment the PR opens, reopens, or gets a new commit — no CLI call required at all.
- Without the App, the next
uploads attachyou run against that PR triggers the same promotion + comment refresh as a side effect. If you have nothing new to add right after opening the PR, runuploads attach --promote(zero file arguments) to promote and refresh explicitly — it exits0even if nothing was staged. Skip auto-promotion on a given call with--no-promote.
"PR already exists" is just the simple case of the same command — same
uploads attach, just pointed at a PR/issue number instead of a branch, and
the comment updates immediately since there's already something to comment
on:
uploads attach ./before.png ./after.png
uploads attach ./flow.gif --issue 45 --repo myorg/myapp
uploads attach ./shot.png --no-comment # stable URLs only, no comment
For a URL you'll hard-code in a PR/issue body (re-uploads overwrite in
place, URL never changes). Putting an image in the description works — that's
the right place for a visual that belongs in the write-up (a before/after in
"what it looks like"). You don't have to. Wait until the PR exists and use
--pr, so the body gets the stable pull/<n>/ key from the start. Don't
paste a branch-staged URL into the description and later rewrite it to
pull/<n>/ just to "correct" it.
uploads put ./after.png --pr 123 --alt "Dashboard after" --width 700
For a durable public link to share anywhere (Slack, docs, a teammate):
uploads put ./demo.gif --format url
When you do embed, use the returned markdown (or embedUrl) — the
no-cache host so overwrites propagate. Don't hand-build storage URLs.
Comment briefly disappeared? Don't panic-repost. If the App is installed
and subscribed to the issue_comment event, a deleted or edited-out managed
comment self-heals automatically on the next webhook delivery — no need to
run comment/attach again just to bring it back.
Removed the wrong screenshots? delete the object(s) and re-run
comment (or the hosted comment tool) to re-sync. Once the last attachment
is gone the comment is rewritten in place to a neutral empty state — it stays
on the PR (a later upload repopulates it) rather than leaving stale entries
that point at deleted files.
Bot comment not showing up at all? The managed comment needs a repo↔workspace binding (normally created implicitly by the first comment/promote call, or by installing the GitHub App). If a comment you expected doesn't appear, check the binding first:
uploads github link --status
That's read-only and shows the current binding (or that the repo is
unbound) without claiming anything. If the CLI reports not_authorized
instead, the repo is already bound to a different workspace — it
won't fall back to posting via your own gh auth in that case. The fix is
uploads github unlink --repo owner/name from the owning workspace, or
asking an operator to reassign the binding; switching to the workspace that
already owns it also works.
Curate, don't dump. The comment inlines up to 16 images; anything past
that collapses into a <details> overflow list. Name and pick shots
meaningfully (before.png/after.png, not capture-1..capture-40) rather
than attaching every incidental screenshot from a long session — a curated
handful of milestones reads better than a dumped folder.
Step 3 — Embed well
Meaningful alt text, always (
--alt).Constrain width on large shots with
--width(emits sized<img>).Before/after reads best side by side:
| Before | After | | ------------------------------------ | ----------------------------------- | | <img width="380" src="…/before.png"> | <img width="380" src="…/after.png"> |Motion: GitHub markdown won't autoplay MP4 URLs — prefer a GIF, or a still image that links to the video URL.
Write bodies to a file and use
gh pr edit --body-file/gh issue comment --body-filerather than inline HEREDOCs.
Setup and escalation
- No shell? Use the hosted MCP. Do not install the CLI.
- CLI missing on a machine with a shell?
npm install --global @buildinternet/uploads - Not authenticated on the CLI?
uploads login(one-time, opens a browser), thenuploads doctorto verify. Hosted MCP uses OAuth on first tool call. - Everything deeper — flags, key layouts, MCP tool contracts, metadata and
search, galleries, config defaults, output formats, exit codes — lives in
the uploads-cli skill and
uploads <command> --help.
Cautions
- Uploads are public and effectively permanent until deleted. GitHub repo
visibility is not an access control, and
gh/<owner>/<repo>/pull/<num>/…keys are predictable. Never upload secrets, tokens, or customer PII — crop/redact first. - Private repos get an unguessable URL automatically. When the uploads
GitHub App can see a target repo is private, attachments key under
gh/private/<id>/…instead — no flag needed. The id is durable, not access-controlled: anyone who gets the URL can read it until you rotate it (uploads github rotate-prefix). See docs/private-attachments.md for the full model.
Need to point at something in the screenshot?
Boxes, arrows, labels, freeform strokes, and redaction (solid for secrets
caught in a capture — blurred text can be recoverable) get baked onto a
screenshot with uploads screenshot --annotate or uploads annotate — see the annotate-screenshots
skill for the spec format and workflow, then come back here to attach.