ai.instantclips由 instantclips.ai 發布官方發布者v1.10.1更新於 Nov 6, 2026
Create branded vertical video ads from ecommerce product URLs or product photos.
InstantClips 把電商商品頁面或商品照片變成直式短影音廣告，讓助理匯入商品、起草方案並生成影片。
Web executable，透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。
把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式，或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南，先向你展示發布者、端點和登入方式，經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。
閱讀 https://sourceweft.com/mcp/ai.instantclips%2Finstantclips/install.md，並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前，先告訴我它的發布者、端點和登入方式，等我確認。不要把 API 金鑰寫進命令或我會提交的檔案；問我金鑰應從哪裡取得。把它新增到你客戶端的 mcpServers 設定中。
{
"mcpServers": {
"instantclips": {
"type": "http",
"url": "https://app.instantclips.ai/mcp"
}
}
}InstantClips turns an e-commerce product into short-form vertical video for TikTok, Instagram Reels and Stories. It runs a hosted MCP server, so Claude Code, Codex, Cursor, VS Code, the Claude app, ChatGPT or any other MCP client can do what the web app does: import a product, draft the plan, and generate the video.
This is not a text-to-video generator. InstantClips reads the product page — photos, price, details — and builds the ad from what is actually there. The plan is written first and shown to you; the video follows the plan. That is why it is cheap enough to run across a catalogue, and why the result is the product you sell rather than a guess at it.
Free to start: the welcome credits cover the first video, and there is no card to enter. After that, one-time credit packs, or an Agency membership for anyone running several brands. See pricing.
The product server stays hosted. This repository contains its connection guide, registry metadata, example HTTP client and a small open-source stdio adapter for clients that cannot connect to a remote server directly. The adapter answers initialization, ping and tool discovery from a generated snapshot, then sends authenticated tool calls to the hosted endpoint. The hosted server remains the source of truth; the product implementation is not duplicated here.
Paste one address into your assistant and sign in when it asks. That is the whole setup. The first time it calls, the server sends you to sign in to InstantClips and approve the assistant; there are no keys to copy. Every assistant you approve is listed under Authorized apps in settings, where you can disconnect it.
The connection is to your own account: same brands, products, credits and plan limits as the web app. Signing in creates an account if you do not have one, with credits for the first video.
Opening the endpoint in a browser returns the setup page rather than a protocol error, with one-click install buttons for Cursor and VS Code.
Connect to the hosted endpoint directly whenever your client supports Streamable HTTP; it takes you through sign-in on first use. Use a token and the stdio adapter (under "No browser?" below) only for scripts and automated runners that cannot open a sign-in page.
Paste this into Claude Code, Codex or any agent that runs commands. It adds InstantClips itself, then tells you how to sign in:
The instructions it follows are app.instantclips.ai/mcp.md: the commands below, one section per client.
Then run /mcp inside Claude Code and choose instantclips to sign in. --scope user adds it to
every project; without it, Claude Code adds a server only to the folder you run the command in.
It adds InstantClips to ~/.codex/config.toml, which the Codex CLI, the Codex IDE extension and the
ChatGPT desktop app all read, and opens your browser to sign in. codex mcp login instantclips
starts the sign-in again.
One-click install buttons are on the setup page. They open the app, add InstantClips, and sign you in on first use.
https://app.instantclips.ai/mcp and select Add.On a Team or Enterprise plan, an owner adds it first under Organization settings › Connectors, and each member then connects it in Customize › Connectors.
https://app.instantclips.ai/mcp with OAuth, then sign in to InstantClips when asked.On a Business, Enterprise or Edu workspace, an admin creates the app under Workspace settings › Apps › Create and publishes it for the workspace.
In ChatGPT, attach the product photos to the conversation and ask for the video: the plugin takes
attachments directly (image_files on create_product_from_images, add_image_files on
update_product), so photos need no adapter and no token there.
Menu names are the apps' own as of September 2026. If they have moved, Claude's custom connector guide and OpenAI's developer mode guide have the current ones.
OpenClaw, Hermes, or an agent you wrote yourself: point it at the address over Streamable HTTP. The server announces its sign-in flow the standard way, so a client that follows the spec needs nothing else. Anything that cannot open a sign-in page uses a token instead, below.
Scripts, CI jobs and agents that cannot open a sign-in page authenticate with a long-lived token instead. Mint one at app.instantclips.ai/settings#ai-access. It gives full access to your account, so keep it out of anything you commit.
With a token, the same clients look like this:
To keep the token out of the Codex file, swap the header for bearer_token_env_var = "INSTANTCLIPS_TOKEN"
and export it in your shell instead. Any other client sends an Authorization: Bearer header.
Nothing on the wire is InstantClips-specific. The chat apps always have a browser to sign in with,
so they use the steps above rather than a token.
The instantclips-mcp npm package is a thin stdio-to-HTTPS adapter. It serves initialization and
tool discovery locally for a fast, credential-free cold start, then reads the token from the
environment and sends tool calls to InstantClips:
For an automated connectivity check, including the live tool names:
The token is accepted only through INSTANTCLIPS_TOKEN, never as a command-line argument, so it
does not appear in the process list. It is required for tool calls, but not for initialize,
ping, or tools/list. Node.js 20 or newer is required.
Photos on this machine: the adapter can upload them itself, so a video can be made from files
that never left your laptop. Through the adapter, create_product_from_images takes image_paths
(paths to the files, up to 9, 8 MB each) in place of image_urls, and update_product takes
add_image_paths. The files are packaged and posted straight to InstantClips as the product's
photos — nothing is downloaded from anywhere. Drag the files into a client that hands the
assistant their paths (Claude Code, Cursor, VS Code, terminal agents) and say what you want made.
Send image_paths on its own: with hosted image_urls in the same call the adapter refuses
rather than dropping the URLs; add hosted photos afterwards with update_product.
Five to begin with. Swap in a link or a product name.
The workflow, in order:
import_product_from_url for a store page, or create_product_from_images when
there is no page to read.get_product until import_status is "imported".ideas, as the website does. The
user's pick, or their own words, goes in update_video_direction's creator_note, which writes
the plan; redraft_video_direction with no idea picks one for them. The plan comes back as
text for the user to review, and update_video_direction changes anything in it.generate_video, passing expected_credit_cost: the cost the user was told, as
get_product reported it. A mismatch is refused without charging.get_video until it is done, then give the user download_url, the video's
own page and its public share link.Another video for the same product is the same flow on that product: editing or redrafting opens the
next video's draft, and generate_video with no draft renders another take of the last plan. A
generated video itself cannot be changed. Every get_product response carries next_step: what to
do now. Once the user has watched a video, rate_video records their verdict; send_feedback sends
the team a note about the tools themselves, with the user's OK.
Brands work the same way: list_brands, create_brand, set_product_brand. Every video is
drafted in a brand's voice, so an import whose storefront matches no existing brand stops and asks
rather than guessing.
Each tool's exact parameters are published by the hosted server. The generated
manifest/instantclips-mcp.json snapshot lets stdio clients and
registries inspect those same schemas without a credential. Maintainers refresh it with
npm run sync:manifest, no token needed, since the hosted server lists its tools without signing in;
npm run check:manifest fails, naming what drifted, when the committed snapshot differs from the
live server. Run python example.py tools below when you specifically
want to print the live schemas over HTTP.
Importing a product, drafting the plan and editing it are all free. generate_video is the
only tool that spends credits, and it requires your explicit go-ahead — the tools report the cost
first, and generate_video takes that number back as expected_credit_cost, refusing a launch whose
cost has changed. An agent cannot quietly run up a bill. See pricing.
A dependency-free MCP client — Python 3.9+, standard library only, no pip install. It
authenticates with a token, since a script has no browser to sign in with.
tools is the one to run first: it prints the real parameter names and types for every tool, which
is what you need before scripting the workflow above.
server.json is this server's entry in the official MCP registry,
which the other directories ingest from. One ai.instantclips/instantclips entry carries both the
hosted endpoint in remotes and the stdio adapter in packages, so a host can choose the transport
it supports without creating two identities for the same tool surface.
The npm package's mcpName must exactly match that registry name. The repository link points to the
open-source adapter; the hosted product implementation is not in this repository.
The ai.instantclips namespace is the reverse-DNS of the domain, which requires publishing under
DNS or HTTP domain auth rather than GitHub auth. Authenticating with GitHub instead would force the
entry into io.github.instantstudioai/... and give up the branded namespace.
Publish the npm package first, then re-publish this same registry entry with
mcp-publisher publish after bumping its version. Domain authentication keeps the branded
ai.instantclips namespace; do not replace it with an io.github.* name. The signing key stays out
of the repository — .gitignore covers *.pem, and a committed private key is a published one.
npm is published from GitHub Actions, not from a laptop. Commit the version bump (package.json,
server.json, and the pins in test/shim.test.js), then push a tag named for it
(git tag v1.7.2 && git push origin main v1.7.2); .github/workflows/publish.yml runs the tests
and publishes through npm's trusted publishing, with no npm token and no browser confirmation. The job stops if the tag does not match package.json's version. The package's
trusted publisher on npmjs.com names this repository and publish.yml, so renaming the workflow
file breaks publishing.
A new version can take several minutes to appear on npm, and the registry refuses an entry whose
npm version it cannot find ("version … was not found"; nothing changes). Run the mcp-publisher
step once https://registry.npmjs.org/instantclips-mcp/<version> answers.
Log in and publish in one command — the registry token expires five minutes after login — and the
domain is verified over HTTP, not DNS: instantclips.ai/.well-known/mcp-registry-auth on the marketing site
serves this key's public half (v=MCPv1; k=ed25519; p=…), and there is no TXT record, so
login dns fails with "no MCP public key found".
test/shim.test.js pins the snapshot's server version and server.json's version; a release
moves both.
glama.json is the separate, Glama-specific file that claims the listing there. A server under an
organisation rather than a personal account can only be claimed with that file present.
It carries ownership only. In Glama's Dockerfile form, use build steps
["npm install --omit=dev"], CMD arguments ["node", "./bin/instantclips-mcp.js"], and any dummy
value for the required INSTANTCLIPS_TOKEN placeholder. Glama's initialization and tool-quality
checks use the bundled manifest and never transmit that placeholder upstream. Do not put a real
account token into a third-party build sandbox.
MIT — see LICENSE. The license covers this repository's contents; use of the hosted service is governed by the Terms of Service.
來源：README.md，提交 7574974
List the account's brands
The account's brands, the plan's brand limit, and `account`: `credits_remaining`, `top_up_url`, and `watermarked: true` on a free account. Call it before answering `brand_decision_required`, so the user chooses a brand by name.
無需輸入參數。
List the account's products
The account's products, newest first, including ones made on the website. Use it to find a product_id you don't have, then `get_product`. Carries `account`, as `list_brands` does.
limitquerybrand_idImport a product from its page URL
Start a product from its page URL; returns at once, so poll `get_product`. It detects the brand, and an external AI service reads the page into facts and ideas. A store locked to another account is refused and reported to support.
url必填brand_idCreate a product from photos
Start a product from photos when it has no page (`import_product_from_url` when it does). Photos carry no brand: ask the user which, or pass `brand_id`. An external AI service then reads the photos; poll `get_product`.
name必填brand_idimage_urlsdescriptionimage_filescreator_noteGet a product and its video direction
Read a product: import status, brand decision, photos, the video direction with its settings and `credit_cost`, and its videos. Poll every 20-30 s while it imports or drafts. `next_step` says what to do now.
product_id必填Create a brand for a product
Create a brand for a product waiting on `brand_decision_required` and attach it; then offer its ideas. Ask the user first: it takes a brand slot. A photos product needs `name` from the user, and the voice, market and keywords they give.
namevoicekeywordsproduct_id必填target_marketAttach a product to an existing brand
Attach a product waiting on `brand_decision_required` to an existing brand (`list_brands`), once the user confirms it belongs there; then offer its ideas (`get_product`). `replace_identity` rewrites that brand for every product on it.
brand_id必填product_id必填replace_identityEdit a brand's identity
Edit a brand's `name`, `voice`, `target_market` or `keywords` (a new list replaces the old); only the fields you pass change. Future drafts use them; `redraft_video_direction` rewrites a direction already drafted.
namevoicebrand_id必填keywordstarget_marketEdit a product's facts, link and photos
Edit a product's name, description, price and posted `link_url`, and add or remove photos; only the fields you pass change. The direction is drafted from these facts, so `redraft_video_direction` after changing them.
namepricelink_urlproduct_id必填descriptionadd_image_urlsadd_image_filesremove_image_idsEdit the video direction and render settings
Edit the next video's `creative_direction`, its idea (`creator_note`), settings and photos; only the fields you pass change. Settings are parameters, not direction text. A new idea, length or ratio rewrites the direction: poll `get_product`.
ratioproduct_id必填resolutioncreator_notetarget_marketduration_secondscreative_directionselected_image_idsDraft a new creative direction
Draft the next video's direction through an external AI service: with none and no idea yet, from one of the product's ideas we choose; else another take in its format. It overwrites edits: confirm first. Poll `get_product`.
product_id必填Generate the video (spends credits)
Render the video. SPENDS THE USER'S CREDITS: ask first, with the direction, settings, cost and any watermark, and pass the cost as `expected_credit_cost`. External services render it; it gets a public share page and may send account email.
product_id必填expected_credit_cost必填Check a video's render status
Check one video's render. `next_step` says what to tell the user and when to poll again; follow it, since it knows how long this render has run. Mention a watermark only when `watermarked` appears. Done or failed: `launched_with` is its plan.
video_request_id必填Rate a finished video
Record the user's own verdict on a finished video they have watched: `reaction`, and optional `issues` and `comment` in their words. Never rate for them, and ask at most once. A second call replaces the rating.
issuescommentreaction必填video_request_id必填Tell the InstantClips team about a problem with these tools
Send the InstantClips team a note about these tools: one that failed or surprised, or something the user wanted that no tool does. Show the user the note and send it only with their OK. Not for support or ratings.
message必填已滿足適用檢查項的 100%
該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。
Set me up with InstantClips by following the instructions at https://app.instantclips.ai/mcp.mdclaude mcp add --transport http --scope user instantclips https://app.instantclips.ai/mcpcodex mcp add instantclips --url https://app.instantclips.ai/mcp# Claude Codeclaude mcp add --transport http --scope user instantclips https://app.instantclips.ai/mcp --header "Authorization: Bearer YOUR_TOKEN"# Codex, in ~/.codex/config.toml[mcp_servers.instantclips]url = "https://app.instantclips.ai/mcp"http_headers = { Authorization = "Bearer YOUR_TOKEN" }{ "mcpServers": { "instantclips": { "command": "npx", "args": ["-y", "instantclips-mcp"], "env": { "INSTANTCLIPS_TOKEN": "YOUR_TOKEN" } } }}INSTANTCLIPS_TOKEN="your-token" npx -y instantclips-mcp --check --jsonexport INSTANTCLIPS_TOKEN="your-token"
python example.py tools # every tool, with its live input schemapython example.py call list_brands '{}' # call one tool with JSON argumentsmcp-publisher login http --domain instantclips.ai \ --private-key "$(openssl pkey -in key.pem -text -noout | awk '/priv:/{f=1;next} /pub:/{f=0} f' | tr -d ' :\n')" \ && mcp-publisher publish