Adding MCP problem type basic sampler infrastructure#1628
Conversation
Introduces a new problem domain for blackbox testing of MCP (Model Context Protocol) servers, following the same architecture as the existing GraphQL and RPC problem domains. New components under core/src/main/kotlin/.../problem/mcp/: - McpAction (abstract base), McpToolCallAction (tools/call), McpResourceReadAction (resources/read), McpInputParam, McpUriParam - McpIndividual: chromosome wrapping sequences of MCP actions - McpCallResult: per-action execution record storing isError flag - client/McpClient (interface), client/HttpMcpClient (JSON-RPC 2.0 over HTTP with pagination), client/McpDTOs - service/McpSampler: @PostConstruct discovery via tools/list, resources/list, resources/templates/list; gene tree building from JSON Schema; random and smart (ad-hoc) sampling - service/McpFitness (abstract base), service/McpBlackBoxFitness: per-action scoring using isError flag and successful reads as signals - service/McpBlackBoxModule: Guice bindings mirroring GraphQLBlackBoxModule Unit tests cover action construction, individual copy/mutation, and HttpMcpClient JSON-RPC parsing via WireMock stubs.
- Add ProblemType.MCP (experimental) to EMConfig enum - Add bbTargetUrl validation for MCP black-box mode - Route ProblemType.MCP to McpBlackBoxModule in Main.init() - Add getAlgorithmKeyMcp() supporting RANDOM, MIO, MOSA, WTS, SMARTS - Dispatch MCP algorithm key in Main.run() - Bind NoTestCaseWriter in McpBlackBoxModule EvoMaster can now be invoked with: --blackBox true --problemType MCP --bbTargetUrl <url>
Connection failures during @PostConstruct were surfacing as an uncaught InvocationTargetException labelled as an EvoMaster bug. Wrap the listTools() call so a failed connection to the MCP server produces a clean, user-readable SutProblemException instead.
Servers may return 400/404/405 for MCP methods they don't support (e.g., resources/list on a tools-only server). Previously this caused an IOException from HttpURLConnection.inputStream on non-2xx responses, crashing the sampler during initialization. Extracted openConnection() helper and made post() return null on 400/404/405, which callers treat as an empty/error result.
* Setting up MCP Problem Type Skeleton * remove db related size from MCP individual * Fix for resource read action id Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> * feedback * adjust resource read action parameters --------- Co-authored-by: guido-rodriguez_sfemu <guido.rodriguez@salesforce.com> Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
c7cc742 to
d4c0c34
Compare
| /** | ||
| * MCP protocol version negotiated during the initialize handshake. | ||
| */ | ||
| const val PROTOCOL_VERSION = "2025-11-25" |
There was a problem hiding this comment.
Just as a question, could this be a parameter for when fuzzing an MCP API? Or are protocol versions not backward compatible?
| mapOf( | ||
| "protocolVersion" to "2024-11-05", | ||
| "capabilities" to emptyMap<String, Any>(), | ||
| "clientInfo" to mapOf("name" to "EvoMaster", "version" to "1.0.0") |
| ?: return McpToolResult(isError = true) | ||
| val result = response["result"] as? Map<String, Any?> ?: return McpToolResult(isError = true) | ||
| val rawContent = result["content"] as? List<Map<String, Any?>> ?: emptyList() | ||
| val content = rawContent.map { c -> |
There was a problem hiding this comment.
Favour readability, this could be extracted to a getMcpContent method
| val rawContent = result["content"] as? List<Map<String, Any?>> ?: emptyList() | ||
| val content = rawContent.map { c -> | ||
| val type = c["type"] as? String | ||
| ?: throw IllegalStateException("tools/call response content item missing required 'type' field") |
There was a problem hiding this comment.
Can we add more information on which tool call is actually causing this error? If we're fuzzing to test an API, it's best if we provide the most information to our users. Applicable to the other exceptions being thrown
| import javax.ws.rs.client.Entity | ||
| import javax.ws.rs.core.MediaType | ||
| import javax.ws.rs.core.Response | ||
| import kotlin.String |
There was a problem hiding this comment.
Are all this imports needed?
| val result = response["result"] as? Map<String, Any?> ?: return McpResourceResult() | ||
| val rawContents = result["contents"] as? List<Map<String, Any?>> ?: emptyList() | ||
| val contents = rawContents.map { c -> | ||
| McpContent( |
There was a problem hiding this comment.
How come this McpContent can be created without type and we throw an exception for the one in callTool? It's not a good practice to use the same object to represent different things. Either this should have a type assigned, or it should be a different object
| /** Content item within a tool or resource response. */ | ||
| data class McpContent( | ||
| val type: String, | ||
| val type: String? = null, |
There was a problem hiding this comment.
Everything is nullable? Why? I fear we might hide issues due to this flexibility
| } | ||
|
|
||
| // Discover tools | ||
| val tools = mcpClient.listTools() |
There was a problem hiding this comment.
Maybe extract this and subsequent discoveries to their own methods. That way we have a more concise and declarative initialize method
| return ind | ||
| } | ||
|
|
||
| val n = randomness.nextInt(1, getMaxTestSizeDuringSampler()) |
There was a problem hiding this comment.
What's n? Choose a better name
Summary
Adding the basic search infrastructure for blackbox fuzzing of Model Context Protocol (MCP) servers. The approach is: connect to the target server, discover its capabilities via the standard MCP handshake, automatically construct a mutable input representation (gene tree) for each tool and resource from their JSON Schema definitions, then use a search algorithm to evolve a sequences of calls that maximise coverage; reaching each capability and obtaining a non-error response
Components
MCP client (HttpMcpClient, McpClient, McpDTOs)
An HTTP-based client that implements the MCP JSON-RPC protocol: initialization handshake, tool invocation (tools/call), static resource reads (resources/read), and capability discovery (tools/list, resources/list, resources/templates/list).
Action model (McpAction, McpToolCallAction, McpResourceReadAction)
Two concrete action types mirror the two MCP interaction patterns: calling a named tool with structured arguments, and reading a resource by URI (including template URIs with mutable parameters).
Individual (McpIndividual)
A test case represented as an ordered sequence of McpActions, integrating with EvoMaster's existing EnterpriseIndividual and mutation infrastructure.
Sampler (McpSampler)
On startup, connects to the target MCP server, performs the handshake, and discovers all tools and resources. Each tool's JSON Schema input definition is converted into an ObjectGene tree used by the search engine to mutate arguments. Pre-built single-action individuals ensure every capability is exercised at least once before random search begins.
Fitness function (McpBlackBoxFitness)
Evaluates individuals by executing their actions sequentially and scoring coverage targets based on observable protocol signals: 1.0 for a successful tool call or resource read, 0.5 for a server-reported error (partial credit keeps the search gradient useful). Exceptions break the action sequence early.
McpBlackBoxModule
Wires all the above into EvoMaster's dependency injection container and registers the problem type in the main entry point.
Follow Up
client calls is needed.
(explicit JSON Schema format → type switch → name-based heuristics) to generate valid values for these fields, matching the approach used by the REST builder.