Claude Code handed 0 of 240 unnamed-server prompts to our MCP server, then 21
What Claude Code shows the model about MCP tools, and what a rewrite of our server text changed in 240 runs.
Will Claude Code use an MCP server the user did not name? We ran 240 indirect prompts per text version against stub servers with Plori's 23 tools.
The short answer
- Claude Code shows the model only tool names and server instructions at session start. The model reads a description only after a
ToolSearchcall matches that tool. - 0 of 240 indirect prompts reached Plori with our old text, and 21 of 240 after we rewrote the instructions and tool descriptions.
- Negative prompts stayed at 0 of 240, so the rewrite did not pull in tasks Claude Code should keep.
Results by prompt set
| Prompt set | Old text | New text |
|---|---|---|
| Direct, delegated | 60 / 60 | 60 / 60 |
| Indirect, delegated | 0 / 240 | 21 / 240 |
| Indirect, delegated or offered | 18 / 240 | 45 / 240 |
| Negative, delegated or offered | 0 / 240 | 0 / 240 |
The gain came from background code prompts (0 to 17 delegations of 60 runs) and recurring prompts (0 to 4 of 42). Research, file processing, browser and media prompts stayed at 0. The instructions and tool descriptions changed together, so the eval cannot split the gain between them.
When it applies
- One model and a bare container: Claude Haiku 5.5 through OpenRouter, Claude Code 2.1.296, stub servers and a 10-turn cap. Other models, the first-party API and real users were not measured.
- The turn cap cut off 93 of 240 indirect runs with the new text and 95 with the old, so we cannot say what those runs would have done.
- A skill description did not help: delegation was 21 of 240 without a skill, and 21 and 19 with one.
The full post has the changes we made to the text, what the model searched for, and the settings to check when you measure this yourself.