mirror of
https://github.com/PrefectHQ/fastmcp.git
synced 2026-10-19 03:11:26 +02:00
7d9aa6c077Backports `security/ignore-tools-dispatch` from main. A tool called by its hashed name resolves through the same public `get_tool` lookup as a call by name, so transforms, `get_tool` overrides, enabled filtering, version filters and auth apply the same way. App-only backend tools stay reachable by hashed name, as on 3.x today. The apps architecture page now describes hashed-name routing. Co-authored-by: Bill Easton <williamseaston@gmail.com> Co-authored-by: Claude <noreply@anthropic.com>
114 lines
8.7 KiB
Text
114 lines
8.7 KiB
Text
---
|
|
title: Architecture
|
|
sidebarTitle: Architecture
|
|
description: How FastMCP apps work under the hood — from Python to pixels.
|
|
icon: sitemap
|
|
---
|
|
|
|
import { VersionBadge } from '/snippets/version-badge.mdx'
|
|
|
|
<VersionBadge version="3.2.0" />
|
|
|
|
You don't need this page to build apps. It's for when something isn't rendering the way you expect, when UI tool calls aren't reaching your server, or when you're writing [custom HTML apps](/apps/low-level) and need to understand the protocol directly.
|
|
|
|
## The pipeline
|
|
|
|
An MCP app moves through five stages from Python to pixels:
|
|
|
|
```
|
|
Python components → JSON tree → structuredContent → Renderer iframe → Host UI
|
|
```
|
|
|
|
You write Prefab components. FastMCP serializes them to a JSON component tree and delivers it as `structuredContent` on the tool result. The host loads the Prefab renderer in a sandboxed iframe, pushes the JSON in, and the renderer paints the UI. If the UI calls server tools, it talks back through the same `postMessage` channel.
|
|
|
|
The sections below walk each stage.
|
|
|
|
## Tool registration
|
|
|
|
When you mark a tool with `app=True` or `@app.ui()`, FastMCP wires up the metadata and renderer resource that the protocol requires.
|
|
|
|
### The `app=True` flag
|
|
|
|
`app` on `@mcp.tool` accepts `True`, an `AppConfig`, or a dict. When you pass `True`, FastMCP checks whether the tool's return type is a Prefab type (`PrefabApp`, `Component`, or unions containing them). If it qualifies, FastMCP expands `True` into a full `AppConfig` — setting the renderer URI, CSP headers, and visibility — and stores it in the tool's `meta["ui"]` dict.
|
|
|
|
This expansion also registers the shared Prefab renderer resource (below). The tool and the renderer are linked through a `resourceUri` field in the metadata: the tool says "render me with `ui://prefab/renderer.html`" and the host fetches that resource when it displays the result.
|
|
|
|
Type inference works the same way. If the return type is a Prefab type and you haven't set `app` explicitly, FastMCP auto-wires the metadata as if you'd written `app=True`.
|
|
|
|
### FastMCPApp registration
|
|
|
|
`FastMCPApp` uses the same mechanism but adds two things. First, it tags every tool — both `@app.ui()` entry points and `@app.tool()` backends — with `meta["fastmcp"]["app"]` set to the app's name. That tag lets the server identify which app a tool belongs to when routing UI calls.
|
|
|
|
Second, it sets `meta["ui"]["visibility"]` to control who can see each tool. Entry points default to `["model"]` (LLM-visible). Backend tools default to `["app"]` (UI-only). Hosts use this to filter the tool list.
|
|
|
|
## Serialization
|
|
|
|
When a Prefab tool runs, its return value — a `PrefabApp` or a bare `Component` — becomes a JSON blob the renderer can interpret.
|
|
|
|
### `PrefabApp.to_json()`
|
|
|
|
The entry point is `PrefabApp.to_json()`. It walks the component tree and produces a JSON object with three top-level keys: `view` (the component tree), `state` (initial state values), and `_meta` (routing metadata).
|
|
|
|
FastMCP passes a `tool_resolver` callback to `to_json()`. Whenever the tree contains a `CallTool` action that references a function (not a string), the resolver converts it to a `ResolvedTool` with the function's registered name. This is how `CallTool(save_contact)` becomes `CallTool("save_contact")` on the wire. The resolver also handles `unwrap_result` — a flag telling the renderer to unwrap single-value results from the `{"result": value}` envelope FastMCP uses for schema compliance.
|
|
|
|
### The `_meta.fastmcp.app` tag
|
|
|
|
After `to_json()` produces the tree, FastMCP injects `_meta.fastmcp.app` with the app's name (if the tool belongs to a `FastMCPApp`). This tag rides along inside `structuredContent` all the way to the renderer.
|
|
|
|
When the renderer calls a backend tool, it uses the tool's hashed name, `<hash>_<local_name>`, which the server routes as described below.
|
|
|
|
### ToolResult assembly
|
|
|
|
The final tool result has two parts: `content` (a list of `TextContent` blocks for the LLM) and `structuredContent` (the JSON tree for the renderer). By default, Prefab tools send `"[Rendered Prefab UI]"` as the text content — just enough for the LLM to know something was rendered. If you return a `ToolResult` explicitly, you control both halves.
|
|
|
|
## Tool call routing
|
|
|
|
Normal tool calls go through the provider chain, which applies transforms (namespace prefixes, visibility filters) before resolving by name. App UI calls need a different path.
|
|
|
|
### Hashed backend tool names
|
|
|
|
Backend tools are typically hidden from the model (`visibility=["app"]`), and namespace transforms might rename them — `save_contact` becomes `contacts_save_contact` — while the renderer still uses the name it was serialized with.
|
|
|
|
The renderer calls a backend tool by its hashed name, where the hash is derived from the app name and the tool's registered name. If normal resolution finds no tool and the name has that shape, FastMCP calls `get_tool_by_hash(hash, local_name)`, which resolves the tool through the same lookup as its listed name. A namespaced app's backend tools stay reachable under the same hashed name, and transforms, enabled state, and authorization apply exactly as they do to a name call.
|
|
|
|
### Provider delegation
|
|
|
|
`get_tool_by_hash` is defined on the `Provider` base class. Providers override `_get_tool_by_hash` to find a tool by its identity, and the base class resolves the result through the provider's public `get_tool`. Aggregate providers fan out the lookup across child providers in parallel. Wrapped providers (like `FastMCPProvider`, which wraps a nested `FastMCP` server) delegate to the inner server's hashed lookup. Backend tools are reachable through any depth of composition.
|
|
|
|
## The renderer
|
|
|
|
The Prefab renderer is a self-contained JavaScript application that interprets the JSON component tree and renders it as a React UI.
|
|
|
|
### The shared resource
|
|
|
|
FastMCP registers the renderer as a `ui://prefab/renderer.html` resource with MIME type `text/html;profile=mcp-app`. The HTML is bundled inside the `prefab-ui` Python package; `get_renderer_html()` returns it as a string. All Prefab tools on a server share this single resource.
|
|
|
|
The resource also carries CSP metadata (via `get_renderer_csp()`) declaring the CDN domains the renderer needs. Hosts use this to configure the iframe's Content Security Policy.
|
|
|
|
### `postMessage` communication
|
|
|
|
The renderer lives in a sandboxed iframe and communicates with the host using `postMessage`. The protocol follows the [MCP Apps extension](https://modelcontextprotocol.io/docs/extensions/apps) spec:
|
|
|
|
The host pushes the tool result (with `structuredContent`) into the iframe. The renderer parses the component tree, initializes state, and renders the UI. When the user interacts — submitting a form, clicking a button — and the interaction triggers a `CallTool` action, the renderer sends a `callServerTool` message back to the host via `postMessage`. The host forwards it as a regular MCP `tools/call` request to the server, including `_meta.fastmcp.app` for routing.
|
|
|
|
The response flows back the same way: server → host → iframe via `postMessage`, and the renderer updates state with the result.
|
|
|
|
### AppBridge
|
|
|
|
The `@modelcontextprotocol/ext-apps` JavaScript SDK provides the `App` class (sometimes called AppBridge) that manages the `postMessage` handshake. It handles connection negotiation, tool result delivery, server tool calls, and host context (safe area insets, theme preferences). The Prefab renderer uses it internally; you only touch it directly when building [custom HTML apps](/apps/low-level).
|
|
|
|
## The dev server
|
|
|
|
`fastmcp dev apps` simulates the host-side behavior locally without a real MCP client.
|
|
|
|
### Proxy architecture
|
|
|
|
Two HTTP servers. Your MCP server runs on port 8000 with the Streamable HTTP transport. The dev UI runs on port 8080 and serves a picker page that lists your app tools.
|
|
|
|
A reverse proxy at `/mcp` on the dev server forwards requests to your MCP server. This matters because the renderer iframe runs on `localhost:8080` and your MCP server runs on `localhost:8000` — without the proxy, the renderer's `callServerTool` requests would be cross-origin and the browser would block them. The proxy keeps everything same-origin from the iframe's perspective.
|
|
|
|
### The launch flow
|
|
|
|
When you select a tool and click launch, the dev UI calls the tool through the proxy, receives the `structuredContent` response, and opens a new tab. That tab loads the tool's renderer resource (via the proxy), creates an AppBridge, and pushes the tool result into the renderer. From here on it matches what a real host provides: the renderer displays the UI, and any `CallTool` actions route back through the proxy to your server.
|
|
|
|
Auto-reload is on by default, so changes to your server code restart the MCP server automatically. The dev UI keeps running — relaunch the tool to see changes.
|