How to soak-test your MCP server before AI agents do it for you
Most MCP servers get tested the same way: connect one client, call a few tools by hand, ship it. Then real agents show up. Dozens of sessions at once, three tool calls in parallel, sessions that never get closed. The failures that follow don't look like crashes. They look like memory that creeps up for hours, a p95 that slowly doubles, or Session not found errors that only appear once you run two…
This article explains how to perform a soak test on MCP servers using an open-source tool called mcpload. The primary difference between a load test and a soak test is that the former focuses on how much stress a server can handle, while the latter examines whether the server remains healthy over an extended period.
The soak test follows a three-phase process: warm-up, steady load, and cool-down. During the warm-up phase, the server's startup growth, including caches and connection pools, is allowed. The steady load phase maintains a constant stream of agent sessions for 30 minutes, while the cool-down phase involves no load for 5 minutes, allowing memory to return to its baseline.
Mcpload simulates real agents by performing a series of tool calls, including initialization, listing tools, and executing tool calls in parallel. The test scenario provided in the article involves running 3,680 sessions over 30 minutes, with two new agent sessions starting every second.
The article includes a real-world example using the official MCP reference server, server-everything. On a laptop with limited resources, the server handled 3,780 sessions, generating 49,227 requests and experiencing no errors. Memory usage remained stable at around 136 MiB, with no significant increase during the soak test. The p95 latency remained steady at approximately 21 ms for all three tools over the entire test period.
The author also notes that throughput scales linearly up to 40 concurrent agents, with no errors observed. However, at 80 concurrent agents, latency increased significantly, although there were still zero errors. The results highlight the importance of testing a server's limits before users encounter any issues.
The remaining parts of the article discuss additional testing scenarios, such as running the test behind a load balancer, testing for session not found errors, handling session state, and running the tests in a Continuous Integration (CI) environment. The author also mentions running the same test across other official MCP SDKs and encourages readers to share their findings or suggest improvements to the mcpload tool.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
