What is the maximum number of tools an mcp server can have for Claude Desktop?

[yadgire7]

Pre-submission Checklist

• I have checked that this question would not be more appropriate as an issue in a specific repository
• I have searched existing discussions and documentation for answers

Question Category

• Protocol Specification
• SDK Usage
• Server Implementation
• General Implementation
• Documentation
• Other

Your Question

I connected a remote MCP server with 206 tools to Claude Desktop app. After connection, I can see only 100 tools. Is there a limit on maximum number of tools a server can have?

[saen-ai]

arround 50 more then that it the quality will be not that much good according to my experience

[olaservo]

Hi @yadgire7 , I don't think that Claude Desktop support is monitoring these discussions - I would take a look at the support options here instead: https://support.claude.com/en/articles/9015913-how-to-get-support

I'd also echo @saen-ai 's comment that loading that many tools at once will typically have a negative impact on performance (and burn through your Claude usage faster).

[isheng-eqi]

There is no hard protocol-level limit in MCP, but there are practical limits driven by Claude's token context window. Here is what is actually happening when you see only 100 out of 206 tools.

The real limit: token context window

Claude Desktop (and Claude Code) has a maximum context window of 200k tokens. Every tool registered with the MCP server consumes tokens before the conversation even starts — the tool's name, description, and full inputSchema are injected into the system prompt.

Here is rough math:

Component	Per-tool cost
Tool name (avg)	~5 tokens
Description	~20-50 tokens
inputSchema (JSON)	~100-500 tokens
Total per tool	~150-550 tokens

For 206 tools, even conservatively: 206 × 200 = 41,200 tokens consumed just by tool definitions. If your tools have rich schemas, it could be 80k-100k+ tokens.

Why you see exactly 100

Claude likely applies internal truncation when the tool list exceeds a threshold. 100 is a common batch size in many LLM API implementations. This is NOT an MCP protocol limit — it is Claude's client-side optimization to keep the system prompt within reasonable bounds.

The maintainer reply suggesting you contact Claude support is partially correct — the limit lives in Claude's client, not in MCP.

Practical strategies for 200+ tool servers

1. Toolset grouping (best approach)

The GitHub MCP server solved this exact problem. It has 91 tools but groups them into toolsets:

github-mcp-server:
  --toolsets code_security,actions,issues

Only the tools in enabled toolsets are registered with Claude, keeping the context footprint small. If your server supports per-toolset registration, use it. If it does not, consider adding it — this is the standard pattern.

2. Multiple server instances by domain

Instead of one server with 206 tools, split into domain-specific servers:

charging-server:    tools for EV charging (50 tools)
weather-server:     tools for weather data (50 tools)
analytics-server:   tools for analysis (56 tools)
billing-server:     tools for billing (50 tools)

Each server then has a manageable tool count, and Claude connects to only the servers relevant to the current task.

3. Shorter descriptions

Tool descriptions do not need to be documentation. A one-line description that a human and LLM can both understand is sufficient. Cut verbose descriptions to reduce token cost.

4. Lazy/dynamic tool registration

Advanced: register only a "router" tool with Claude that accepts a domain parameter, and have the server dynamically expose domain-specific tools based on the current conversation context. This requires custom server logic but scales to thousands of tools.

Bottom line

The 100-tool limit is not an MCP issue — it is Claude protecting its context window. The fix is architectural: use toolsets (if your server supports them), split into domain servers, or implement lazy registration. Do not try to force 206 tools into one connection — the LLM quality degrades long before the token limit is hit.