Small models via ANTHROPIC_BASE_URL fail to make tool calls due to excessive tool definitions (259 tools sent)

Description

When using Claude Code with a local model via ANTHROPIC_BASE_URL (e.g., Ollama), the model receives ALL configured tool definitions (built-in + all MCP server tools) in every API request. For projects with multiple MCP servers, this can mean 200+ tool definitions, which overwhelms small models and prevents them from generating tool_use responses.

Environment

• Claude Code: 2.1.42
• Ollama: 0.15.6
• Model: qwen3:8b (8B parameter)
• Node.js: v20.19.6
• OS: Ubuntu 22.04 (Linux 6.8.0-90-generic)
• MCP servers configured: 13

The Problem

The model CAN use tools — proven via direct API call:

# Direct Ollama API call with 1 tool = PERFECT tool_use response
curl -s http://localhost:11434/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: ollama" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "qwen3:8b",
    "max_tokens": 256,
    "tools": [{
      "name": "bash",
      "description": "Run a bash command",
      "input_schema": {
        "type": "object",
        "properties": {"command": {"type": "string"}},
        "required": ["command"]
      }
    }],
    "messages": [{"role": "user", "content": "Use the bash tool to run: echo hello_world"}]
  }'

Response (correct tool_use):

{
  "content": [
    {"type": "thinking", "thinking": "The user wants me to use the bash tool..."},
    {"type": "tool_use", "id": "call_esd2pde4", "name": "bash", "input": {"command": "echo hello_world"}}
  ],
  "stop_reason": "tool_use"
}

But through Claude Code, the same model outputs plain text instead of tool_use blocks, because it receives 259 tool definitions:

17 built-in tools (Task, TaskOutput, Bash, Glob, Grep, Read, Edit, Write, etc.)
+ 242 MCP server tools (from 13 configured MCP servers in .mcp.json)
= 259 total tools

How I Diagnosed This

1. HTTP Proxy interception: Created a transparent proxy between Claude Code and Ollama to capture the actual API requests.

1. Proxy log showed: {"num_tools": 259, "model": "qwen3:8b"}

1. Debug log analysis (--debug "api"): No tool_use events in the response. The model generated text instead of structured tool calls.

Secondary Issues Found

1. Token counting uses hardcoded haiku model (404 errors)

Claude Code uses claude-haiku-4-5-20251001 for countTokensWithFallback. When ANTHROPIC_BASE_URL points to Ollama, this model doesn't exist, causing hundreds of 404 errors:

countTokensWithFallback: haiku fallback failed: 404
{"error":{"type":"not_found_error","message":"model 'claude-haiku-4-5-20251001' not found"}}

Each tool definition triggers a separate token counting attempt, leading to ~9+ errors per MCP server.

2. --strict-mcp-config partially helps but isn't sufficient

Using --strict-mcp-config with a minimal MCP config successfully reduces MCP tools, but:

• The 17 built-in tools are always included (no way to filter them)
• The large CLAUDE.md system prompt is still loaded
• Combined, this still overwhelms 7-8B parameter models

3. MCP config loading performance

With --strict-mcp-config, MCP config loading took 943 seconds (15+ minutes):

[STARTUP] Loading MCP configs...
[STARTUP] MCP configs loaded in 943498ms

Reproduction Steps

1. Install Ollama, pull qwen3:8b
2. Create a project with multiple MCP servers in .mcp.json
3. Run:

``bash
ANTHROPIC_BASE_URL="http://localhost:11434" \
ANTHROPIC_API_KEY="ollama" \
claude --model "qwen3:8b" --print \
"Use the Bash tool to run: echo hello_world"
``

1. Observe: model outputs text instead of making tool calls
2. Compare: direct curl to localhost:11434/v1/messages with 1 tool definition produces correct tool_use response

Expected Behavior

When using ANTHROPIC_BASE_URL with a non-Anthropic model:

1. There should be a way to limit which tools are sent (e.g., --include-tools "Bash,Read,Write" or --max-tools N)
2. Token counting should gracefully skip when haiku isn't available (not error per-tool)
3. MCP config loading should not take 15+ minutes

Suggested Fix

A --small-model mode or --include-tools flag:

claude --model "qwen3:8b" \
  --print \
  --include-tools "Bash,Read,Write" \
  --strict-mcp-config \
  --mcp-config minimal.json \
  "Use Bash to run: echo hello"

Impact

This effectively prevents using Claude Code with small local models for any agentic task. The tool flooding makes --model with Ollama models decorative-only (text generation, no tool use). Users who want to leverage Claude Code's infrastructure (MCP servers, hooks, permissions) with local models cannot do so.