How to soak-test your MCP server before AI agents do it for you

说明与用途

mcpload 持久性测试能发现 MCP 服务器在并发长时间运行时出现的内存泄漏、会话错误和延迟问题。

问题

MCP servers are often tested with a single client and a few manual tool calls, which misses issues that arise under concurrent, long‑running sessions such as memory leaks, session‑not‑found errors, and latency spikes.

结果

The soak test on a laptop‑based server with 2 CPUs and 1 GiB memory processed 3,780 sessions and 49,227 requests with no errors, memory remained flat (~136 MiB), and p95 latency stayed around 21 ms. Scaling to 80 concurrent agents increased latency sharply but still produced zero errors, indicating the server’s healthy limits.

工作流

• Install mcpload from GitHub releases
• Start the MCP server in Docker with fixed resources
• Run a one‑minute smoke test to verify basic functionality
• Execute a 38‑minute soak test with warm‑up, steady load, and cool‑down phases
• Analyze memory and latency metrics from the soak output
• Optionally test load balancer scenarios, never‑closed sessions, and burst traffic
• Integrate mcpload into CI with a GitHub Action for automated testing

证据

Most MCP servers get tested the same way: connect one client, call a few tools by hand, ship it.

来源

A leak is flagged only if memory grows steadily during the load window (slope above a limit, with a good linear fit) or doesn’t come back near its baseline during cool‑down.

来源

3,780 sessions, 49,227 requests, 0 errors

来源

Memory flat: ~136 MiB, growing 0.12 MiB/min against a 1 MiB/min limit, and back near baseline after cool‑down

来源

p95 latency steady at ~21 ms for all three tools over the full 30 minutes

来源