AgentGrid.io (by Anima) MCP Server認証済み注目

@Anima

リポジトリホームページドキュメント

基本情報

トランスポート

Streamable HTTP

認証方式

OAuth

カテゴリ

生産性

作者

Anima

投稿者

matan-schmidt

AgentGrid.io (by Anima) MCP Server について

Anima による、人間と AI エージェントのための共有ワークスペース。あらゆるエージェントからライブアプリ、ドキュメント、ファイルを作成・バージョン管理・公開・共有でき、エージェントごとのアクセス制御を備えています。あらゆる種類のエージェントに対応：Grok Bot、Muse AI、OpenAI Dots、Claude、Claude Code、Claude Design、ChatGPT、Cpdex、Cursor、OpenClaw、Hermens など。

#agent workspace#multi-agent #mcp #ai agents #claude #claude code #codex #cursor#figma to code#publish web app#share artifacts#collaboration

接続情報

https://api.agentgrid.io/v1/mcp

セットアップ

claude mcp add agentgrid-io --transport http https://api.agentgrid.io/v1/mcp

AgentGrid.io (by Anima) MCP Server のツール

Get the manifest.json file describing the design system documentation structure. Returns a JSON object mapping file/folder paths to their metadata (description, type, optional tags/category).

Fetch multiple documentation files by their paths. Returns a JSON object mapping file paths to their content. Use this after getting the manifest to retrieve specific files like README.md, COMPONENTS.md, TROUBLESHOOTING.md or component documentation.

Check the build status of an EXISTING artifact (from artifact-create). Read-only. Generation is asynchronous, so this is how you find out it finished. Call it with { sessionId, wait: true } and it WAITS for you: the call returns as soon as status is 'ready' or 'failed', or after about 45 seconds still 'generating'. Then, you can call it again. This tool is meant to be called right after artifact-create, before replying to the user, and keep getting called until status is no longer 'generating'. ONLY when you want an instant snapshot, this tool can be called without wait (or false); the idea is to never spin on wait: false. Requires sessionId — the id returned by artifact-create (or the last path segment of an artifact URL like https://app.agentgrid.io/artifacts/<sessionId> or https://dev.animaapp.com/chat/<sessionId>). **Returns:** { success, sessionId, status: 'generating' | 'ready' | 'failed', progress (0–100, while generating), name, artifactUrl (generating/ready only), playgroundUrl (app and knowledge artifacts only, generating/ready only), previewUrl (app, knowledge and markdown artifacts only, generating/ready only), error (when failed), nextStep (when a wait returned still generating) }

Inspect the files of an EXISTING artifact: list them, search their text, read them, or review the commit history. Works with no shell, no git and no network access, so prefer it whenever you cannot run git. Pass the sessionId (the last path segment of a .../chat/<sessionId> URL). A canvasDraft field, when authorized, reports an uncommitted Canvas draft. Use action draft to inspect its projection overlay explicitly; ordinary reads remain committed files. Draft access requires write permission. Typical flow to change something: action "search" to find the file, action "read" for the files you will edit, then artifact-edit. Every response includes "revision", the artifact's current commit — pass it back as artifact-edit's baseRevision. Use action "read" with a "revision" to see how a file looked at an earlier commit (this is how you undo something). Assets — images, fonts, media — are readable too: action "read" answers with size, mime and oid under "asset": true whenever a file's bytes are not text, instead of content. That check is on the bytes, so it is the authority: an SVG reads as text, and a file with an unfamiliar extension may still come back as an asset. When a search comes back empty, check "notSearched": "assets" counts binary files a text query can never match, and "excluded" counts files under node_modules, dist or build — re-run with includeExcluded true to search those. A "tree" listing also maps any path artifact-edit cannot change under "unwritable". For generation or build progress use artifact-status instead; this shows the commit log, not build state.

List the workspaces you can access and what you may do in each. Takes no parameters. Workspace ids are opaque: this names the ones you can reach, and a workspace-list_artifacts row carries the id its artifact is filed in. **Returns:** { success: true, workspaces: [{ workspaceId, name, isGeneral, capabilities }] }. `capabilities` is a subset of read, write, share, publish; you need `write` to create or change an artifact in that workspace. `isGeneral` marks the team's General workspace. An empty list means you have been granted no workspace, not that the team has none: ask the user for access.

List the artifacts you can read, most recently updated first — across every workspace you can access, or in one with `workspaceId`. Each row's `sessionId` is what artifact-explore, artifact-edit and artifact-get_git_token take. **Returns:** { success: true, artifacts: [{ name, type: app | knowledge | markdown | asset, sessionId, workspaceId, workspaceName, updatedAt }], truncated, nextSteps }. Rows cover only what you can access, and the listing is capped (`truncated: true` means more exist): report what you found, never a total, and never say the team has no artifacts.

Get git access to an EXISTING artifact, for local development with a real git client. Use this when you HAVE a working shell with git and outbound network access AND the job suits a local checkout — a large refactor, running or testing the project, branches, or history rewriting. For ordinary reading and editing of an artifact’s files, prefer artifact-explore and artifact-edit: they need no shell, no git and no network of your own, and they change the same repository. An artifact is a real git repository. Pass the sessionId — the id of an existing artifact (e.g. the last path segment of a .../chat/<sessionId> URL, like "mr25vsjppVtbMx") — and this returns a gitRemoteUrl plus the authenticated principal’s commitAuthor. After cloning, apply the returned `git config user.name` and `user.email` instructions before committing; then edit files, commit, and `git push` — pushing updates the live artifact. The gitRemoteUrl holds a short-lived access token scoped to this one artifact (read-only or read-write, depending on your access). Tokens CANNOT be renewed: on a "token expired" git error, call this tool again for a fresh gitRemoteUrl and run `git remote set-url origin <new gitRemoteUrl>`, then retry. If a git command instead fails because the host cannot resolve or reach the server (DNS, proxy or firewall errors), do NOT retry it — that environment has no route to the git remote, so use artifact-edit instead. Treat the gitRemoteUrl as a secret. To rename an artifact or change its visibility, use artifact-update_metadata.

Read the change requests humans have addressed to YOU on their artifacts. A review is a batch of comments a reviewer wrote together and sent as one, each pinned to a place in the artifact. Call this when a human says they left you comments or asked for changes, and at the start of work on an artifact you have been reviewed on before. Every comment carries an anchor. For an app artifact that is `dom`: `sourceFile` with `sourceStart`/`sourceEnd` points straight at the code when it resolved, and `instanceCount` above 1 warns that the markup is SHARED — changing the component changes every instance, so change the one instance unless the reviewer meant all of them. For a document it is `markdownRange`: find the text by SEARCHING the markdown for `quote` (with `prefix`/`suffix` to tell repeats apart), never by the offsets, which any edit above them invalidates; `headingPath` says which section it is in, and `endQuote` with `spansBlocks` marks a comment covering several blocks. `isLatest` false means the artifact moved since the review was written — pull and judge whether the comments still apply before acting. You see the comments addressed to you, plus any addressed to nobody, which arrive marked `unassigned`: a review may hold comments for several agents, and another agent's are not yours to act on or report on. An unassigned comment is anyone's to take: reply to say you are taking it, which announces the work but does not reserve it — the comment stays open to every agent, and another may be working the same one. A comment addressed to PEOPLE and not to you never reaches you at all. A comment may name several parties, agents and people together. `alsoAsked` lists everyone else it names, so you can see a colleague is on it and not redo their work. Any of them may resolve it, and resolving closes it FOR ALL OF THEM. The mentions in the body may scope parts of the request — "@you fix the header, @someone else the footer" — and that scoping is in the words, not in the fields. So when you have done only your part, REPLY rather than resolve: resolving would close the others' half too, and they would never learn it was dropped. A comment may point at nothing in particular — that is the reviewer's note about the artifact as a whole, and it carries no anchor. Each comment carries its own `status` and `replies`, the conversation so far. READ THE REPLIES before acting: they may answer a question you asked, and they may also add or change what is being asked for, since a reviewer can keep talking after sending. The comment body is where the request starts, not necessarily where it ends. To resolve, copy the review's `resolveTrailer` into your commit message — it is the exact lines that close these comments, one per comment. Drop the lines for any comment the commit does not address. Use review-resolve when no commit will, and review-reply to say something without closing it.

Close ONE review comment WITHOUT a commit — for what a commit message cannot carry: a question answered, or a deliberate decision not to make the change, explained in `note`. When a commit does address the comment, prefer the trailer `AgentGrid-Resolves: <reviewId>#<index>` in its message: that records which commit resolved it, which this cannot. Closing ends the exchange — if you are asking the reviewer something rather than answering them, use review-reply instead, which leaves the comment open for them to come back to. You may close a comment addressed to you, or one marked `unassigned`. Safe to retry — a comment already closed answers resolved: false rather than failing.

Say something back on ONE review comment and LEAVE IT OPEN — to ask what the reviewer meant, to report what you found, or to say you are blocked. Use this rather than review-resolve whenever the exchange is not finished: resolving closes it, and the reviewer would have to open a new comment to answer you. The reviewer sees your message in the artifact, and their answer comes back on the same comment — call review-list again to read it. They may not reply immediately, so do not wait on it. You may reply to a comment addressed to you, or to one marked `unassigned`. A closed comment takes no more replies — it answers replied: false, because nobody is reading that thread any more. Use `mentionUserIds` to bring a PERSON into the comment when you need a human decision, an approval, or information you cannot get yourself: pass the `id` of anyone in the artifact's `mentionablePeople` from review-list — they need not already be on the comment. Each person named is emailed, so name somebody only when the exchange genuinely needs them. You cannot mention another agent.

概要

What is AgentGrid.io?

AgentGrid.io is a shared workspace by Anima where humans and AI agents build, host, publish, and share live web apps, docs, and files. Connect any MCP-capable agent, including Claude, Claude Code, ChatGPT, Codex, Cursor, and VS Code, and everything it creates lands in one governed hub your team can see, edit, and share.

What your agent can do

Build an app from a prompt, a website, or a Figma design: describe it, clone a URL, or hand over Figma frames, and get back a running app at a live URL.
Host code you already have: import a project or start an empty repo, and push to it.
Edit any artifact over git: every artifact is a real git repository, so you can clone, commit, and push.
Publish to a live public URL: and take it down again.
Generate design-aware code into your own codebase: Figma frames to files in your repo, matched to your stack.

Everything is an artifact: one git repo in your team's workspace.

Quick start

Server URL: https://api.agentgrid.io/v1/mcp · Transport: streamable HTTP · Auth: OAuth

Claude Code plugin

/plugin marketplace add AnimaApp/mcp-server-guide
/plugin install anima@mcp-server-guide

This configures the MCP server and installs the skill. Authenticate when prompted.

Claude Code (manual)

claude mcp add --transport http anima https://api.agentgrid.io/v1/mcp

Restart Claude Code, run /mcp, select anima, and authenticate in the browser.

OpenAI Codex

codex mcp add anima --url https://api.agentgrid.io/v1/mcp
codex skill install AnimaApp/mcp-server-guide/skills/anima

Cursor, VS Code, and any other client

{
  "mcpServers": {
    "anima": {
      "url": "https://api.agentgrid.io/v1/mcp"
    }
  }
}

No MCP? Use the CLI

npx @animaapp/cli@latest login
npx @animaapp/cli@latest create -t p2c -p "SaaS dashboard with sidebar and analytics"

Example prompts

"Build me a SaaS analytics dashboard with a sidebar, KPI cards, and an activity feed."
"Clone stripe.com/payments into an app."
"Turn this Figma frame into a working site: https://figma.com/design/..."
"The header on this artifact is broken on mobile, fix it."
"Implement this Figma design in my project."
"Publish it."

Best practices

Describe intent, not CSS. AgentGrid.io is design-aware. Say what it's for, who it's for, and the 3 to 5 features that matter, rather than hex values and pixel sizes.

Be specific when implementing. "Implement the login form using our existing Button and Input components from src/components/ui" beats "implement this design".

Break down large designs. Header first, then the card grid, then the footer.

Troubleshooting

MCP not recognized: confirm the server is configured in your tool and that authentication completed.
Authentication issues: remove and re-add the server, clear cookies, restart your client.
More: troubleshooting reference

Resources

よくある質問

AgentGrid.io (by Anima) のリモート MCP サーバーとは何ですか?

AgentGrid.io (by Anima) のリモート MCP サーバーは、https://api.agentgrid.io/v1/mcp にホストされた Model Context Protocol エンドポイントです。AI アシスタントはローカルに何もインストール・実行することなく接続できます。

AgentGrid.io (by Anima) の MCP サーバーにはどう接続しますか?

エンドポイント https://api.agentgrid.io/v1/mcp を Claude Code、Cursor、VS Code など任意の MCP 対応クライアントに追加してください。このページのセットアップ手順で各クライアントを一度に設定できます。

AgentGrid.io (by Anima) の MCP サーバーは認証が必要ですか?

はい。AgentGrid.io (by Anima) は OAuth を使用します。初回接続時に MCP クライアントがブラウザを開いてサインイン・認可を行い、以降はその認証情報を再利用します。

AgentGrid.io (by Anima) の MCP サーバーはどのトランスポートを使用しますか?

AgentGrid.io (by Anima) は Streamable HTTP エンドポイントを公開しており、これはリモート MCP サーバーで広く使われ、主要な MCP クライアントがサポートするトランスポートです。

他のリモート MCP サーバー

AOI — Environmental Intelligence

AOI is a commercial environmental-intelligence gateway that gives AI agents programmatic access to 38 real-world intelligence capabilities — climate risk, satellite monitoring, methane detection, water & soil, vegetation

42min

Official 42min MCP server: let Claude, ChatGPT, Cursor or any MCP client manage your meeting scheduling. Check availability, find free slots, book, cancel and reschedule meetings, manage recurring series, secret booking

2pr.io

Write, schedule and publish LinkedIn posts in your own voice, reply to comments and read your analytics, all from your AI assistant. Remote MCP server with OAuth sign-in.

Adobe Marketing Agent

Marketing campaign and audience insights from Adobe

AdisInsight

Pharmaceutical drug & clinical trial intelligence