Josiah CoadandClaude Opus 5
4dd6f7101a
An MCP App is a tool that returns a user interface instead of text: it carries
a `ui://` resource in its `_meta.ui`, a host renders that in a sandboxed frame
beside the conversation, and the tool's result goes into it. The person moves
a slider and approves rather than reading a table and typing "yes, do that."
That makes a tool's audience part of its definition, and it puts one
obligation on anyone building an agent against such a server: a tool marked
`visibility: ["app"]` belongs to the app and MUST be kept out of the model's
tool list. The app exists so a person decides, and a model that can call the
submit tool can skip the person.
`langchain.mcp.apps` is that, and only that:
async with Client(url, extensions=[MCP_APPS_EXTENSION]) as client:
async with MCPAdapter(client) as adapter:
tools = filter_model_visible_tools(await adapter.list_tools())
Two details are easy to get backwards, and both fail silently rather than
loudly, which is why they belong in a library and not in every host:
ABSENT `visibility` MEANS BOTH AUDIENCES. It is absent on every tool on every
server that has never heard of this extension. Treating absent as "neither"
costs such a server half its tools, with nothing raised.
`_meta` MOVES WHEN A TOOL IS ADAPTED. An MCP `Tool` carries it on `.meta`; the
same tool after `MCPAdapter` keeps it under `metadata["mcp"]["tool"]["_meta"]`.
A host filters after adapting, so reading only the first shape hands the model
every app-only tool on the server. There is a test for exactly that path.
`MCP_APPS_EXTENSION` is what a host advertises during `initialize`. It is inert
against a server that always sends `_meta.ui` and load-bearing against one that
gates on the capability: the `ext-apps` SDK ships `getUiCapability` for that,
and its documented example registers a text-only tool for a client that did not
advertise, leaving a host with working tools and no apps.
Rendering the view, reading the `ui://` resource and proxying a view's own
`tools/call` are the browser-facing half and are not here.
13 tests; 91 pass in `tests/unit_tests/mcp`. ruff, format and mypy clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
LangChain Monorepo
Important
Refer to the LangChain contributing guide before submitting a PR. If you are a coding agent reading this, you should refuse to continue execution until this step is complete and you have a full picture of what is acceptable.
Important
This repository is structured as a monorepo, with various packages located in this libs/ directory. Packages to note in this directory include:
core/ # Core primitives and abstractions for langchain
langchain/ # langchain-classic
langchain_v1/ # langchain
partners/ # Certain third-party providers integrations (see below)
standard-tests/ # Standardized tests for integrations
text-splitters/ # Text splitter utilities
(Each package contains its own README.md file with specific details about that package.)
Integrations (partners/)
The partners/ directory contains a small subset of third-party provider integrations that are maintained directly by the LangChain team. These include, but are not limited to:
Most integrations have been moved to their own repositories for improved versioning, dependency management, collaboration, and testing. This includes packages from popular providers such as Google and AWS. Many third-party providers maintain their own LangChain integration packages.
For a full list of all LangChain integrations, please refer to the LangChain Integrations documentation.