Stitch MCP skills are a library of Agent Skills Google Labs open-sourced for the Stitch MCP server. Each skill follows the Agent Skills open standard, so coding agents such as Claude Code, Cursor and Gemini CLI can load it and run Stitch design work through a repeatable workflow instead of a one-off prompt. Source: google-labs-code/stitch-skills.
you (plain words)
|
v
+-----------------+ loads +------------------+
| coding agent | <------------------> | skill folder |
| Claude Code / | SKILL.md + refs | (Stitch skills) |
| Cursor / Gemini| +------------------+
+--------+--------+
| MCP tool call
v
+-----------------+ returns +------------------+
| Stitch MCP | -------------------> | layout, screen, |
| server | | design output |
+-----------------+ +------------------+
What Are Stitch MCP Skills?
Stitch MCP skills are reusable instruction packs that teach an AI agent how to use the Stitch MCP server well. A skill is a folder with a description of when to use it and the steps to follow. The agent reads the skill, then calls Stitch tools over MCP. Google Labs published the library as open source.
Two pieces fit together here, and it helps to keep them apart:
- The MCP server is the connection. It exposes tools the agent can call.
- The skill is the knowledge. It says which tools to call, in what order, and what a good result looks like.
That split is the reason this release matters beyond Stitch. Once a design capability lives behind an MCP server, a skill file is all you need to change how an agent approaches it. No new integration, no new plugin code.
Creator OS uses the same pattern. The hosted MCP server at https://mcp.creatoros.ca/mcp exposes about 160 tools for posts, analytics, comments, DMs, automations, webhooks, Skool, WhatsApp, Shopify, the blog and the Ads add-on. Creator OS also ships agent skills, so an AI app can follow a known workflow rather than improvise one. The full tool list is in the MCP docs.
How Stitch MCP Skills Work With Claude Code, Cursor and Gemini CLI
Stitch MCP skills work in any agent that reads Agent Skills and can reach an MCP server, which is why Claude Code, Cursor, Gemini CLI and Antigravity are named in the library. The agent loads the skill into its context, then calls the server’s tools.
The loop looks the same in every one of those clients:
- You describe the design task in plain words.
- The agent matches the request to a skill description and loads that skill.
- The skill points the agent at specific Stitch MCP tools and the order to use them.
- The agent calls the server, gets a result, and either continues or reports back.
For a creator, the practical effect is that you can say “turn this into three mobile screen layouts” and get a repeatable process rather than a different answer every time you phrase it. The skill carries the method. You carry the brief.
7 Stitch Skills That Speed Up a Creator’s Design Workflow
The value of a skill library is coverage, not one clever prompt. For a creator, the fastest wins come from skills that remove the setup steps before every design session.
Here is a seven-part way to think about what a design skill can do for you, mapped to real creator work:
| Skill role | What it saves you | Creator example |
|---|---|---|
| Project bootstrap | Repeating the same context every session | Load brand colours and fonts once |
| Screen and layout generation | Wireframing by hand | Draft a link-in-bio page layout |
| Design critique | Guessing what looks wrong | Ask for a hierarchy and contrast check |
| Component extraction | Redrawing the same button and card | Build a reusable kit for post templates |
| Design to code | Handoff friction | Turn a layout into markup for a landing page |
| Design system checks | Inconsistent spacing and type | Catch off-brand colours before publishing |
| Iteration loops | Restarting from scratch after feedback | Apply one note and re-check the rest |
Notice that only some of these make pixels. Several just make the same request easier to repeat, which is where most creator time actually goes.
Stitch MCP vs One-Off Prompts: Why the Agent Skills Pattern Wins
The Agent Skills pattern wins because the knowledge lives in a file the agent can load on demand, so it survives context changes and gets better as you edit it. A one-off prompt is thrown away after each answer.
Three things break with one-off prompts:
- They drift. Small wording changes produce different processes and different output.
- They do not travel. A prompt that works in one session is not a thing your team can install.
- They cannot be reviewed. You cannot diff a chat message, but you can diff a skill file.
A skill fixes all three. It is versioned, editable and shareable. If a step is wrong, you change one line and every future run improves.
Creator OS applies this idea directly. npx @creatoros/cli@latest init installs 14 Claude Code skills for posting, scheduling, replies, analytics, ads and Skool. creatoros sync updates them without overwriting edits you have made, and saves changes as .new files. It also pulls the campaign builder skills the Creator OS API publishes, currently the Meta, Google, TikTok and LinkedIn campaign builders.
npx @creatoros/cli@latest init
creatoros sync
creatoros validate:post-length --text "your caption here"
The skills are also served as MCP prompts, so they show up in the prompt or slash menu of an MCP-capable AI app. Setup steps are in the Claude Code plugin docs.
How to Set Up Stitch MCP Skills Step by Step
Setup is three moves: connect the MCP server, install the skills, then run one small task to confirm the agent can see both. The install commands and client-specific steps live in the library repository and the Stitch docs. The pattern below is the same one Creator OS uses, so it generalises.
For Creator OS, the steps are concrete:
- In Claude, open Settings, then Connectors, then Add custom connector. Paste
https://mcp.creatoros.ca/mcp, sign in, pick the workspace, and choose read only or read & write. - In Claude Code, add the server with a bearer key:
claude mcp add --transport http creatoros \ https://mcp.creatoros.ca/mcp \ --header "Authorization: Bearer cos_live_..." - In Cursor, Windsurf or VS Code, add an
mcpServersentry with typehttp, the URL and the Authorization header. - Run one read call to confirm the connection, for example a post list. If it returns, the tool path is live.
Two behaviours worth knowing before you hand an agent write access:
- A read-only connection only sees read tools, and the API refuses writes from it.
- Destructive tools such as delete post, delete comment, disconnect account or delete blog article are marked so the AI app asks you first.
Every connection gets its own key, listed under Settings, then Connected apps. That is how you can tell which app did what later. For a walkthrough of the build side, watch KevBuildsApps on YouTube.
Can Stitch MCP Skills Generate Visual Variants for Social Posts?
Yes, if the agent can reach a tool that renders them, but the skill is the workflow, not the renderer. Stitch skills can drive a design session in which an agent produces variants. Turning those variants into posts on real accounts is a separate job, and that job needs a publishing MCP server with write access to your socials.
Creator OS handles the second half. Two features pair well with a design agent:
- Variations takes one finished video and makes up to 20 distinct versions, one per account, so no two accounts post the same file. Each version keeps the content and changes the look, and Claude writes a different hook, caption and hashtags for each. A uniqueness check holds back versions that looks the same. Access is any paid plan, two runs at a time.
- Format reformats one finished video for every network it will post to, applying each platform’s size, length, file size, frame rate and bitrate rules, plus a caption written to each platform’s limits. Each output comes with a ready
POST /v1/postsbody for that platform.
So the split is clean: a design agent explores and generates. Creator OS takes the settled file, makes the per-account and per-platform versions, and publishes them. You can read the post options for each network at the post options docs.
If you want the “one file, many accounts” problem solved first, the agent skills for social media automation post covers it.
What Stitch MCP Skills Mean for Other MCP Servers (LinkedIn, Meta, CapCut)
What Stitch MCP skills mean for other MCP servers is that the barrier to a polished agent workflow is now writing a skill file, not building an integration. Any server that publishes tools can be covered by a skill that explains how to use them.
Creator OS already ships its own campaign builder skills for the paid networks. Those are Creator OS skills, not part of Google’s Stitch library: meta-campaign-builder, google-campaign-builder, tiktok-campaign-builder and linkedin-campaign-builder. They arrive as MCP prompts too, so they surface in the AI app’s menu.
Ads work through the same server and the same review discipline:
- Every create can be a dry run with
validate_only, so the agent proposes and you approve. - Ads are created paused for review.
- Budgets are whole units of the ad account currency.
A worked example, using the CLI:
creatoros ads:create --accountId acc_meta --adAccountId act_123 \
--name "Spring drop" --goal traffic --budget 20 --budgetType daily \
--headline "New prints" --text "Three new pieces, live now." \
--link https://example.com/shop --countries US,CA --validateOnly
That single command shows the shape: the agent knows the goal, the budget and the guardrail because the skill told it, not because you retyped it. The same logic applies to LinkedIn sponsored posts from a Company Page, where LinkedIn reviews every ad and there is no native dry run. Details are in the ads docs.
FAQ
How do I install Stitch MCP skills in Claude Code or Cursor?
Follow the install steps in the google-labs-code/stitch-skills repository, which covers the skill folder layout and the clients it supports. The pattern is to connect the Stitch MCP server first, then place the skill folders where your agent looks for skills. For Creator OS, npx @creatoros/cli@latest init installs 14 Claude Code skills, and creatoros sync updates them later.
Are Stitch MCP skills free and open source?
Google Labs published the stitch-skills library as open source, so the skill files themselves are openly available. You still need your own access to the Stitch MCP server and whichever agent you run. Check the repository licence and the Stitch terms for the current position.
Are the Creator OS campaign builder skills part of the Stitch library?
No. The meta-campaign-builder, google-campaign-builder, tiktok-campaign-builder and linkedin-campaign-builder skills are Creator OS’s own, published by the Creator OS API and pulled in by creatoros sync. The Stitch skills come from Google Labs and cover Stitch design work, not social publishing or ads. Keep the two sets separate when you troubleshoot which skill produced a result.
Can Stitch MCP skills replace Canva or CapCut for social posts?
They serve a different job. Stitch skills guide an agent through design work over MCP, while Canva and CapCut are editing and template apps people use directly. Creator OS complements a design agent by publishing the finished file, making per-account variations and reading the analytics afterwards. Many creators will keep an editor for final polish and let the agent handle volume.
What is the difference between Stitch MCP and a normal MCP server?
The server part is the same idea: a set of tools an AI app can call. The Stitch release adds a skill library on top, so the agent inherits a method for using those tools. Creator OS does the same with about 160 tools plus its own skills, and its hosted server needs nothing installed.
Get started
The fastest way to adopt the Agent Skills pattern for your own accounts is to connect Creator OS and run one small task. The hosted MCP server is at https://mcp.creatoros.ca/mcp, and read-only access is a good first step while you watch what the agent does.
Creator OS plans start at $14.99/month for 11 platforms plus Skool, with the API, the MCP server and the CLI included. Agency plans add a scoped API key per client workspace. Create your account at https://www.creatoros.ca/sign-up, then keep the docs open while you install your first skills. For the visual side of the same idea, the Kokonut UI MCP post shows how an agent moves from components to a real post. For the build walkthroughs, watch KevBuildsApps on YouTube.
Get started for $0
Free trial for new subscribers. Cancel anytime.