Google Stitch MCP

    1

    Generate, iterate and export Google Stitch screens and design systems from any MCP client

    Free

    Security scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Google Stitch MCP

    Example session with this skill installed

    build me three mobile onboarding screens for a habit tracker app — calm, editorial, one accent colour

    • Read your context and instructions
    • Compiled the google stitch mcp

    Project "Habit tracker onboarding" created; its projects/{project} name is kept for the session.
    3 screens generated with deviceType MOBILE; one call timed out, so get_screen was polled until it appeared — nothing was resent.
    Edited "Welcome": headline shortened, illustration moved above the fold, then re-listed to confirm.
    Design system "Calm editorial" created, updated, then applied to the project's screen instances.
    Downloaded 3 screenshots with download_file and listed the local files for review.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Google Stitch is a design generator, and this skill is the playbook for driving it from an agent.

    Stitch exposes 18 MCP tools over Streamable HTTP at https://stitch.googleapis.com/mcp: projects, screens, text-to-screen generation, edits, variants, design systems, direct screen updates and file downloads. Once the endpoint is connected, "build me three mobile onboarding screens for a habit tracker" becomes a sequence of real calls — create a project, generate one focused screen per prompt, list what came back, edit the one that is wrong, branch variants where a direction is still open, apply the design system, then download the files to show.

    It works at the tool level rather than the prompt level: every parameter name, enum and resource-name form comes from the live schema — projects/{project} for get_project, a bare projectId for list_screens, a bare assetId for apply_design_system — and each step's ids are resolved by a previous call, never typed from memory.

    The skill covers:

    • Connection, per client. The exact requirement (a streamable-HTTP MCP server entry named stitch, carrying X-Goog-Api-Key or an Authorization: Bearer token) plus recipes for Claude Code, Cursor, Codex CLI, Gemini CLI, VS Code, Windsurf and the DeepSeek Harness dsh-stitch-mcp plugin — each ending in a proof step, because "connected" is only true when a tool call answers.
    • The real 18-tool catalogue — camelCase parameters, required/optional flags, enums, resource-name forms and MCP annotations — grouped by job, with the call sequences that actually produce usable screens and the "which id form does this call want" table that prevents most failures.
    • Five ordered workflows: idea to first screens, iterating a single screen, exploring variants, applying a design system (including the DESIGN.md path), and auditing or handing off existing work.
    • Failure diagnosis for the signatures people hit: UND_ERR_CONNECT_TIMEOUT, handshake 4xx/5xx, the server/discover probe that can end a connection for good, calls that fail for a missing credential, and a per-origin proxy switch that stays inert.

    It needs a Google account with Stitch access and one credential. There is no helper process, no local port and no installer beyond the MCP client itself — the endpoint is contacted directly over HTTPS, which is why the same skill serves a Cursor user and a DeepSeek Harness user equally well.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for UX & UI Design.

    Stitch returns generated designs, not production code: this skill stops at screen files plus a written spec, and implementing the UI is a separate task. There is no Figma access through this endpoint, and content can only change through the 18 tools Stitch exposes. Generation is slow by design — the server asks for patience and forbids retries — so a session moves at the speed of the design, not the prompt. A network that cannot reach Google needs a proxy before anything works.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 6 days ago

    • Free with an Agensi account

    Accounts
    Google account with Stitch access
    Stitch API key
    orOAuth access token
    Software
    MCP client with Streamable HTTP supportDsh Stitch MCP
    Optional
    Google Cloud SDK, for OAuth tokensHTTP(S) proxy for restricted networksDeepSeek Harness for the plugin path

    Permissions

    Allowed Hosts

    stitch.googleapis.com

    File Scopes

    ~/.dsh/dsh-runtimes/stitch-mcp/
    ~/.dsh/.credentials.yaml
    (read
    only)
    client
    MCP
    config
    files
    (e.g.
    .cursor/mcp.json
    .vscode/mcp.json)
    downloaded
    screen
    in
    the
    working
    directory

    The credential belongs to the user's own MCP client and is never asked for in a file, a repository or a prompt; it is never printed, and the DeepSeek Harness card shows only a mask. Screen downloads write image files into the working directory. Nothing is published or deployed.

    Listed6 days ago

    What's inside

    Frequently Asked Questions

    One credential is required: either an API key sent as X-Goog-Api-Key, created at the Stitch settings page (the full value is shown once), or an OAuth access token sent as Authorization: Bearer, for example the output of gcloud auth print-access-token. The MCP handshake itself needs no credential, which is why the tools can be listed before authentication is configured.

    Yes. Stitch is a standard remote MCP server over Streamable HTTP, so any client that supports remote MCP servers can reach it: Claude Code, Cursor, Codex CLI, Gemini CLI, VS Code, Windsurf. The DeepSeek Harness plugin is one install path among several and only adds a credential card and an opt-in per-origin proxy switch.

    Not as part of this skill. The deliverable is the design: generated and iterated screens, downloaded as files, plus a short written spec of layout, states, copy and spacing. Implementing them in a framework is a separate task with its own skill.

    That is the missing-credential signature. The handshake and the tool listing require no key, so the tools register first; calls then answer with a missing-authentication error until an X-Goog-Api-Key header or a Bearer token is present. The troubleshooting reference maps that and the other predictable failures to their fixes.

    The server's own instruction on generate_screen_from_text, edit_screens and generate_variants is that they can take a few minutes, that a timed-out call must not be retried, and that a connection error does not mean the generation failed. The working pattern is to poll get_screen every 30 seconds up to 10 times, or to re-list screens; resending instead produces duplicates, because these calls are not idempotent.

    Usually not: the endpoint is reached directly and a network that can reach Google needs no configuration. If egress is blocked, set HTTPS_PROXY in the environment that launches the client, or use the DeepSeek Harness card's proxy switch, which routes only the Stitch origin and applies without a restart.

    New in UX & UI Design

    Free