* Propagate CancellationToken request cancellation to remote endpoint
* Refactor existing tests to have a shared base class
* Add cancellation tests
---------
Co-authored-by: Stephen Toub <stoub@microsoft.com>
* Remove McpServerConfig and have McpClientFactory accept IClientTransport instances directly.
* Address feedback and use lists when representing stdio arguments.
* Use "Name" for transport descriptors and "EndpointName" for MCP clients and servers.
* Update code samples.
* Fix source file naming
* Add route pattern parameter to MapMcp
* Add configureOptionsAsync
* Use Kestrel for all in-memory HTTP tests
* Use ApplicationStopping token for SSE responses
* Use RegisterNotificationHandler in MapMcp tests
* Work around SocketsHttpHandler bug where it doesn't call DispseAsync on the stream returned by the ConnectCallback
- We still allow notification handlers to be configured pre-creation, eliminating any race conditions that may result from missed notifications. But now we also allow for post-creation notification handler registration.
- Handlers may now also be removed. This permits temporary registrations that may be created in a scoped manner.
- Registration handling is thread-safe.
- Handlers are now cancelable.
- Exceptions from notification handlers trigger all normal exception handling in the dispatch pipeline.
This also:
- Adds a WithStreamServerTransport as a counterpart to WithStdioServerTransport. The latter was used in several tests in a hacky way, and that's now simplified via the former's existence.
* Move notification handler registrations to capabilities
Currently request handlers are set on the capability objects, but notification handlers are set after construction via an AddNotificationHandler method on the IMcpEndpoint interface. This moves handler specification to be at construction as well. This makes it more consistent with request handlers, simplifies the IMcpEndpoint interface to just be about message sending, and avoids a concurrency bug that could occur if someone tried to add a handler while the endpoint was processing notifications.
* Address more feedback and further cleanup
* Reduce dependencies to 8.x versions of packages
When building for netstandard2.0 or net8.0, the ModelContextProtocol library now only depends on 8.x versions of nuget packages, rather than using the newer 9.x versions. The exception to this are packages that don't have 8.x versions.
This also introduces a net9.0 tfm. When building for net9.0, it'll use 9.x versions of packages.
The PR temporarily removes the Microsoft.Extensions.AI.nupkg reference (the Microsoft.Extensions.AI.Abstractions reference is still very much there). The latest Microsoft.Extensions.AI package on nuget has a System.Text.Json 9.x reference, but the next build to be published drops that down to 8.x as well. Once that build is published, the dependency will be added back, and temporary polyfills (in particular TemporaryAIFunctionFactory) will be deleted.
* Fix StdioClientSessionTransport shutdown
StreamClientSessionTransport's ReadMessageAsync method does await CleanupAsync in its finally. CleanupAsync awaits _readTask. _readTask _is_ ReadMessageAsync. So it's waiting for itself to complete, which will then always timeout. We may want to refactor how this works more in the future, but for now, make sure it doesn't wait for itself.
* Fix dangling requests when server unexpectedly shuts down
If the client makes a request to the server and while waiting for the response the server shuts down, the client hangs. Make sure that when the client's message loops ends due to the server's output it's reading from ends, any pending requests are notified.
* Add net9.0 testing
* Avoid docker port conflicts from test suites running concurrently
* Allow user-defined JsonSerializerOptions for user-defined and inputs and hardcode everything else to the source generator.
* Add validation for user-defined JsonSerializerOptions.
* Remove out of date comment.
* Address feedback and add tests
- This also paves the way for better multi-session support
- We should definitely rethink names for the transport API
- For now, I kept the names similar as possible, so we can
focus on the API shape
Co-authored-by: Stephen Toub <stoub@microsoft.com>
* Overhaul tool handling
- Renames [McpTool{Type}] to [McpServerTool{Type}], in order to distinguish representations of tools on the server from tools on the client.
- Enables [McpServerTool] methods to have arguments injected from DI, as well as a specific set of types from the implementation directly, like IMcpServer.
- Renames WithTools to WithToolsFromAssembly.
- All of the WithToolsXx methods now publish each individual McpServerTool into DI; other code can do so as well. The options setup code gathers all of the tools from DI and combines them into a collection which is then stored in McpServerOptions and used to construct the server.
- The server tools specified via DI as well as manually-provided handlers, using CallToolHandler as a fallback in case the requested tool doesn't exist in the tools collection, and ListToolsHandler augments the information for whatever tools exist.
- The tools are stored in McpServerOptions in a new McpServerToolCollection type, which is a thread-safe container that exposes add/removal notification. Adding/removing tools triggers a change notification that in turn sends a notification to the client about a tools update.
- The ServerOptions are exposed from the server instance.
- Removed cursor-based APIs from McpClientExtensions.
- Changed McpClientExtensions APIs to return `Task<IList<...>` rather than `IAsyncEnumerable<...>`.
I'm sure this will need to evolve further before we're "done", but it's a significant improvement from where we are now. One area I'm not super happy with is the notifying collection; that might work ok for a stdio server, where you can grab the created collection from the server's options and mutate it, but I don't know how well that's going to work for sse servers.
* Make McpServerTool{Collection} extensible
* Fix README.md example
* Add notifications test
* Address feedback
