open-mcp-apps

English | 简体中文

Give your AI a persistent, reusable UI. It builds the app once — you keep it forever.

open-mcp-apps is an open engine built on MCP Apps (ui://, io.modelcontextprotocol/ui) — an extension to the core Model Context Protocol specification, and the first official one, live since 26 January 2026. It gives any MCP-Apps-capable host (Claude Desktop, claude.ai, Codex, ChatGPT, …) three things the extension itself doesn't provide:

1. An app registry the AI can write to. Ask for a UI that doesn't exist — the AI reads the authoring guide, writes a single-file HTML app against a tiny window.oma API, and saves it. From that moment you can open it by name, in this chat and every future one.
2. Persistent, versioned data — separate from the UI. Apps bind to generic collections of items backed by SQLite plus an append-only change_event ledger. Every mutation is an idempotent domain command (command_id) with optimistic concurrency (expected_version). The AI and the human edit the same store — the widget is just a view.
3. A shell runtime so AI-written apps actually work. Serving ui://, the engine wraps the app with the official MCP App bridge, host theming (Claude's design tokens, light/dark), and the window.oma data API. What you write is a view; the protocol, persistence, idempotency and theming are the engine's problem.

Which of those hosts this engine reaches. It runs on your machine and binds 127.0.0.1, so it serves the hosts on that same machine: Claude Desktop, Claude Code, Codex, plus its own browser viewer. A browser host cannot reach a loopback server on your laptop, so claude.ai and ChatGPT web need a remote deployment — today that means the hosted openmcp.app, which runs this same engine for you. Running that remote shape yourself is on the roadmap and not done.

Install

open-mcp-apps runs as a local MCP server. First get it connected to your host (below); then onboarding happens inside the host, separately — that's where the AI builds your first app.

From npm — nothing to clone

If you are comfortable editing your host's config file, point it at the published package and let npx fetch the engine. This path needs only Node 22 — no git, and no checkout for you to keep updated. Paste this into your host's MCP server config:

{
  "mcpServers": {
    "open-mcp-apps": {
      "command": "npx",
      "args": ["-y", "@2nd1st/open-mcp-apps"]
    }
  }
}

The one thing the installer below adds is registering the server into every host it finds. Either way the server seeds the system apps (settings, dashboard, App Store) when it starts, and your data lives in the same fixed per-user store, so you can move between an npx server and a cloned one without migrating anything.

A note on npm: this project publishes under the scoped name @2nd1st/open-mcp-apps. The unscoped open-mcp-apps on the registry is not this project — that name is held by an unrelated package. Check for the @2nd1st/ prefix; the scope is the only thing telling the two apart.

With the installer — one command

Installing needs a shell, so the chat apps (Claude Desktop, Codex) can't install themselves — use one of these instead:

curl -fsSL https://raw.githubusercontent.com/2nd1st/open-mcp-apps/main/install.sh | sh

It opens a short picker to choose which hosts to register into — Claude Desktop, Claude Code, Codex — plus your permission preference. Skip it with -s -- --yes, or target one host with -s -- --host codex.

Where it puts the clone, before you pipe anything into a shell: ~/open-mcp-apps. Set OMA_DIR to put it somewhere else — curl -fsSL <url> | OMA_DIR=~/src/oma sh. Re-running the one-liner updates that same clone in place instead of making a second one. Your apps and data are not in it (they live in the per-user store under Configuration), so the folder is safe to move or delete — and node uninstall.mjs does not delete it for you.

With a coding agent (Claude Code, Codex CLI — they have a shell), paste:

Read https://raw.githubusercontent.com/2nd1st/open-mcp-apps/main/install.md and follow it.

Either way, install.mjs registers the server into each host you pick, idempotently — it never clobbers your other servers, pins a stable node launcher (native SQLite ABI), reports what changed, and cleans up a pre-rename entry if one lingers. Your data lives in a fixed per-user store (not inside the clone), so every host shares the same apps and data.

From a clone — for development

git clone https://github.com/2nd1st/open-mcp-apps && cd open-mcp-apps
npm install
node install.mjs        # same picker as the one-liner above

To wire a clone into a host by hand instead, point it at the checkout — this is the shape install.mjs writes:

{
  "mcpServers": {
    "open-mcp-apps": {
      "command": "node",
      "args": ["/absolute/path/to/open-mcp-apps/src/server.mjs"]
    }
  }
}

Hosted

openmcp.app runs the engine for you. The engine in this repository binds 127.0.0.1 by design, so a self-hosted remote deployment is not a supported shape yet — see Status and roadmap.

Uninstall

node uninstall.mjs unregisters the server from every host it finds — but keeps your data: the shared store stays put, so re-installing later restores every app and all data. It also leaves the installer's clone (~/open-mcp-apps, or wherever OMA_DIR pointed) on disk — nothing here ever deletes that folder, so remove it yourself when you want the checkout gone.

node uninstall.mjs           # unregister from all detected hosts — keeps your data
node uninstall.mjs --purge   # also delete the shared store (apps + data), irreversible
node uninstall.mjs --check   # read-only: show what's registered and what would change

Requirements

• Node 22 or newer, on macOS, Windows or Linux.
• git — only on the installer path. The npx path above needs neither git nor a checkout. The installer checks for both and stops with a message rather than half-installing if either is missing.
• A host that renders ui:// if you want widgets in the conversation. Terminal hosts (Claude Code in a terminal, codex CLI) drive the same data by design and put the UI on a browser screen beside the terminal instead — one that can follow along, showing whatever the AI just opened. The per-host detail is in Host support.
• After installing or updating, fully quit and reopen the host (Cmd-Q, not just closing the window) — it keeps its old server process on the old data until fully quit.

Configuration

Every setting is an environment variable, set in the env block of your host's MCP server entry:

{
  "mcpServers": {
    "open-mcp-apps": {
      "command": "npx",
      "args": ["-y", "@2nd1st/open-mcp-apps"],
      "env": {
        "OMA_VIEWER": "1",
        "PORT": "8787"
      }
    }
  }
}

An app's own tool. Every app opens through open_app. To give one app its own open_<name> tool as well, pin it: Settings → Installed → the app → Own tool. The tool list changes only when you pin, unpin or delete a pinned app, and Claude.ai and ChatGPT show the new tool after you refresh the connector's tool list.

Where your data lives. The whole store is one SQLite file, open-mcp-apps.db, in ~/Library/Application Support/open-mcp-apps/ (macOS), %APPDATA%\open-mcp-apps\ (Windows), or $XDG_DATA_HOME else ~/.local/share/open-mcp-apps/ (Linux). It is outside any clone, which is why every host shares the same apps and data.

First-run permissions. The first few tool calls each show an approval dialog — pick "Always allow". The tool set is small and stable on purpose: read-only tools generally skip approval, and the single open_app tool covers opening every app (including ones the AI creates later) behind that one grant, so nothing new asks again. An OMA_DYNAMIC_TOOLS=1 left in an older entry does nothing; node install.mjs --check reports it as stale, and re-running the installer removes that one key. You can also batch approvals in Settings → Connectors → open-mcp-apps → Tool permissions.

Usage

Start in your host. Restart it after installing. New here? The engine ships one MCP prompt, get_started. A host that surfaces prompts lists it as Get started with open-mcp-apps; hosts that render prompts as slash commands spell it /mcp__open-mcp-apps__get_started. Picking it hands the AI the whole opening move. Not every host surfaces prompts — where yours doesn't, nothing is lost, because the prompt is just a sentence you can say yourself: "I just installed open-mcp-apps — show me how to use it with a couple of examples, and suggest a few apps that fit how I work." Either way it looks at what you already have and what the App Store already offers, asks you a couple of questions, and sets up a first app tailored to you. This step is separate from install and lives in the host. Or just ask directly:

• "make me a board for what I'm juggling right now" → the AI writes it, seeds it, and opens it (persistent)
• "make me a habit tracker" → watch it read the guide, write the app, save it, open it
• close the app, reopen, ask again → everything is still there

The loop

"make me a kanban"
      │
      ▼
list_apps ── exists? ──► open_app {app: "kanban"}     (reuse, instant)
      │ no
      ▼
get_app_guide ──► AI writes HTML ──► save_app
      │
      ▼
open_app {app: "kanban"}  →  rendered inline, themed, persistent — reusable in every future chat

Apps accumulate. Each one is single-purpose and independent — a board, a tracker, a splitter — minted for the task in front of you and kept for the next time you need it.

What it looks like

Apps render inline, in the chat you were already having. Ask for one and the AI writes it:

Come back in another chat — or another host — and it's still there, with your data in it:

The App Store ships ready-made apps, with working previews and one-click install. They live in their own package, @openmcp-app/oma-store, which the engine installs as a dependency — new store apps are contributed there:

Every app above is a single HTML file bound to plain data collections — written with the same window.oma API and authoring guide your AI will use for the apps it builds you.

Multiple widgets in one conversation work fine (habit-streaks + meal-planner side by side).

The browser viewer, and the port it binds

Every install runs a small local web server on http://127.0.0.1:8787. It is how you see your apps outside a chat window — one page per app, the same data your AI is reading — and in a terminal host it is the only way to see them at all, so the AI hands you the link when it builds or opens something.

It starts on its own; OMA_VIEWER and PORT above change that. If the port is already taken by another open-mcp-apps process, that one is already serving the same data and this one just shares its address; if it is taken by something else, you get no viewer and no links rather than a link into a stranger's server.

There is no password on it, and that is deliberate. The listener is hard-wired to 127.0.0.1, so there is no setting that makes it answer from another machine. Any program on your computer that could reach the port can already open the SQLite file directly — a password would be a lock beside an open wall. The one way this reaches the internet is a tunnel you start yourself, which is its own deliberate decision; while a tunnel is up, treat its URL as a secret, because it is currently the only thing standing between the internet and your data.

Host support

Live-tested 2026-07-22 (ChatGPT web 2026-07-28), except cells that carry their own date. A date says when a cell was true, not that it was checked again since.

Everything rides the MCP Apps bridge, so host fixes upstream (e.g. #28912) benefit this project with zero changes.

On Claude Code specifically: it is one product with two surfaces, and only one of them can draw — which is why it takes two rows. In a terminal there is no inline widget surface at all: that is architecture rather than a gap, and it is exactly what a screen beside the terminal is for. The Code surface inside the Claude app has a UI and renders inline; the 2026-08-16 reading there came through the universal open_app.

On Codex specifically: plugins are registered on the web side, so a locally-installed engine is reached as an MCP server, not as a plugin — which is the right path for a self-hosted install anyway. Widget rendering in the ChatGPT desktop app also appears to depend on how you are signed in (we have seen it work under an account sign-in; not yet established under an API key).

A screen beside the terminal

Those two — cells say the chat shows text. They do not say there is no UI. The engine remembers which app was opened last and pushes that pointer to the viewer on the /events frame, and an app can place a region — oma.embed("@live", {into}) — that mounts whatever the AI opened last and swaps itself when the AI opens another. The App Store ships one: install live, open http://127.0.0.1:8787/view/live in a window you then leave alone, and the terminal keeps the conversation while that screen shows the app. The AI opens or writes from the CLI, the screen follows, and you can click, edit and type into it there. Headless CLI use is what it was built for.

The cheapest form of "that screen" is a browser pane in the same tiled workspace — one column over from the agent, in the window you are already working in. Same machine, no tunnel, no second device; most modern terminal setups can put a browser next to a shell, and that is all this needs. A second monitor is the same idea with more desk, and a screen on another device is the same idea again with the caveat below.

The costs are real and worth stating plainly. There is still no widget in the transcript. sendMessage degrades to a notice on a standalone page, exactly as in the Browser viewer row — clicks change data, they do not talk back to the chat. The viewer has to be running (OMA_VIEWER, on by default) with a browser pointed at it; this is not zero-config. And the listener is bound to 127.0.0.1, so "a spare tablet on the wall" means this machine's screen unless you put up the tunnel described above and accept what that section says about it. Inside a chat host the same region deliberately draws a placeholder instead of following anything.

Described from the code as built, not measured on a host — the live-test dates above cover the table, not this section.

Writing an app yourself

The AI is the usual author, but it isn't the only one — its context window shouldn't be the ceiling on what an app can be. Build one in your own editor, with your own bundler, and install it:

node install-app.mjs ./my-app.html              # yours, full trust — same as an AI-authored app
node install-app.mjs ./my-app.html --sandboxed  # untrusted: runs behind the runner, no capabilities
node install-app.mjs --list                     # what's installed, and under whose provenance

# a build pipeline's output: a readable template + its bundle, plus the declaration as its own file
node install-app.mjs ./ui.html --name my-app --manifest ./manifest.json \
  --asset ./dist/app.js --asset ./dist/app.css --update

Two shapes are accepted. One self-contained HTML document — no size cap (keep it lean: data lives in the collection, source is read in windows) — the engine injects a minimal base style, the host's design tokens and window.oma. Or a template plus a bundle: the HTML is a readable mount point that references its own build output (<script type="module" src="https://raw.githubusercontent.com/2nd1st/open-mcp-apps/HEAD/oma-asset:app.js">, <link rel="stylesheet" href="https://github.com/2nd1st/open-mcp-apps/blob/HEAD/oma-asset:app.css">), --asset pushes those files into the app's file plane, and the engine inlines them the moment the document leaves the store — a host iframe cannot reach this machine.

The trade: the AI can no longer iterate on it — your files are the source of truth, you rebuild and re-install. It can still read a single-document app's source; for a template + bundle app it reads the template and the bundle stays yours (edit_app and the AI's save_app refuse with built_outside; re-running the command above is the edit). Either way the app shares your data like any other. Provenance is not overwritable in either direction, so an app installed --sandboxed stays sandboxed until you delete it. What a widget may connect to is the app's own declaration (manifest.csp, relayed to the host per the MCP Apps spec — nothing by default), and an app's functions run engine-side and may fetch — see RUNTIME.md §5.1 and §6.1, and the two type files for what an app sees — @2nd1st/open-mcp-apps/types/window-oma (the window.oma API) and @2nd1st/open-mcp-apps/types/oma-function (the args/api a function body gets) — both resolve through the package exports.

RUNTIME.md is the contract — the window.oma API in both modes, what a sandboxed app can still do, and the traps that only bite authors who aren't the AI. It carries a version, and test/runtime-contract.mjs pins it to the two runtimes' real surfaces, so it can't drift from them silently.

Security model

Trust is tiered by where an app came from — SECURITY.md has the detail.

• Local, full trust: your apps, AI-built apps, the system apps, and everything installed from the App Store — including OMA_APP_STORE_DIRS folders. They run in direct mode with the real window.oma.
• Sandboxed: apps installed with --sandboxed, and apps opened through a share, run behind the runner — a sandboxed srcdoc iframe whose bridge reaches only the app's own collection unless granted more.
• App functions run on your machine and can make network requests. On by default; OMA_FUNCTIONS=0 turns them off.
• The AI can operate the open app (screen_read / screen_act) the way you could, but it cannot confirm a deletion for you.
• Reserved security:* / policy:* config keys are out of reach of generic data writes; the HTTP transport binds to loopback.

Design positions (why it's built this way)

• UI and data persist separately, both versioned. Apps are views; collections are truth; the ledger is history. Swap either without losing the other.
• The AI talks domain commands, never SQL, never raw state. That's what makes human+AI concurrent editing safe (idempotency + optimistic concurrency at the command layer).
• Extension-first. Everything rides the MCP Apps bridge — no host-private APIs. One codebase should serve every host that renders ui://.
• Single-purpose, not composite. Each app owns one scenario and its own collection; the engine mints a new one rather than cramming features into an old one. System apps (settings, dashboard) are the deliberate exception — engine-owned, privileged, allowed to see across collections.

Troubleshooting

Development

npm test                     # every suite below, plus the static invariants and budget checks
node test/server-smoke.mjs   # 461 assertions over real stdio — incl. runtime app creation
node test/http-smoke.mjs     #  85 assertions over the HTTP transport (incl. SSE /events, viewer)
node test/provenance.mjs     #  39 assertions that an app's author — its trust tier — is not overwritable
node test/seed-smoke.mjs     #  30 assertions on the seed pipeline
node test/files-smoke.mjs    #  24 assertions on the per-app file store (dedup, integrity, GC races)

Contributions need nothing signed — MIT in, MIT out (CONTRIBUTING.md).

Status and roadmap

Early v0 — proven end-to-end on Claude Desktop; cross-vendor render + shared store proven on Codex desktop and the browser viewer.

Where it stands:

• engine: registry + shell + generic data commands + ledger
• system apps installed (settings, dashboard, app-store); App Store apps (the @openmcp-app/oma-store package) with live previews, one-click install
• AI app creation loop (guide → save → open)
• in-context onboarding (the get_started prompt → the AI builds a first app for how you work)
• security foundation: trust tiers + sandboxed runner + reserved config keys
• multi-host discovery installer (Claude Desktop · Claude Code · Codex) + shared per-user store
• npx one-command install (@2nd1st/open-mcp-apps on npm)
• self-hosted remote (Streamable HTTP) as a supported shape → claude.ai / ChatGPT / mobile off an engine you run — the transport exists (src/http.mjs) and has been live-tested over HTTPS; what's missing is the hosted story, since the engine binds 127.0.0.1 by design. Those browser hosts already work against the hosted openmcp.app; this box is about doing it yourself
• one-click install with no shell
• app export/import → sharing → community App Store (review + runner sandbox activate here)

License

MIT, for the whole repository — the engine and the system apps in components/ alike (the App Store package is MIT too) (LICENSE · LICENSING.md). Use it, fork it, modify it, embed it, run a modified version as a hosted service; keep the copyright notice with substantial portions you redistribute. That is the whole obligation. Up to v0.5.2 the engine was AGPL-3.0-only under a directory split — see LICENSING.md for what changed and why.

The names open-mcp-apps, openmcp.app, SecondFirst, and 2nd1st, and their logos, are not granted by the license — see TRADEMARKS.md. Fork the code freely; give your fork its own name.

Copyright © 2026 2nd1st.

© 2026 2nd1st