Testing Your MCP with Large Language Models

After completing connect-llm, you're ready to begin testing. While put-into-llm covered basic testing, this article details advanced configuration options visible below the input box.

Model Selection

Switch between different LLMs here. OpenMCP tracks models at the message level, enabling mixed-model testing scenarios.

Can't find your model? See connect-llm to add new models.

System Prompts

Manage and create system prompts through this module:

Prompts are stored in ~/.openmcp/nedb/systemPrompt.db (NeDB format for easy inspection/backup).

Prompt Generation

Invoke server-provided prompt functions, with results inserted as rich text:

Resource Access

Call server resource functions - outputs are inserted as formatted content:

:::warning Data Persistence OpenMCP doesn't manage resource persistence! Empty resources after restart indicate the MCP server lacks storage implementation - this is expected behavior. :::

Parallel Tool Execution

When enabled (default), models may call multiple tools in one response (e.g., three parallel web searches):

Disable for sequential execution.

Note: Some providers (like Gemini) may force-disable this if they don't fully support OpenAI's parallel calling spec.

Temperature Control

Controls output randomness:

Recommended range: 0.6-0.7 for general tasks
Default: 0.6 (balanced creativity/consistency)

Context Window

Determines how many prior messages (default: 20) are sent with each request. Count includes:

User queries
Model responses
Tool calls/results

:::danger Minimum Threshold Values below 20 often break tool call sequencing, causing 400 errors. Start fresh conversations if this occurs. :::

Server Timeout

Default 30-second timeout adjustable via Settings → General (global configuration, in seconds). Increase for long-running operations.

2.1 KiB Raw Blame History