anthropic_tools.py•6.63 kB
# tools/anthropic_tools.py
"""
Specialized Claude Tool - Meta-AI Implementation
This tool demonstrates Claude calling Claude for specialized tasks.
It's "meta-AI" - an AI assistant using another AI assistant as a tool.
Architecture:
- Declarative profile system (CLAUDE_PROFILES dict)
- Single responsibility: orchestrate calls to specialized Claude instances
- Self-registers with shared MCP instance
- Cached client for performance
"""
from core.mcp_instance import mcp
from utils.anthropic_client import get_anthropic_client
# Declarative profile map - keeps logic clean and DRY
# Each profile defines a specialized Claude persona with:
# - system: The system prompt that shapes Claude's behavior
# - description: Human-readable explanation of the profile's purpose
CLAUDE_PROFILES = {
"code_review": {
"system": (
"You are a senior software engineer specializing in code "
"security, performance, and maintainability. Identify "
"vulnerabilities, inefficiencies, and best practice violations "
"clearly and concisely. Focus on actionable feedback."
),
"description": "Performs detailed code and security reviews with actionable feedback."
},
"summarize": {
"system": (
"You are a precise summarizer. Extract key insights, remove "
"redundancy, and produce clear, short summaries. Focus on "
"the most important information and maintain accuracy."
),
"description": "Summarizes text accurately and briefly, preserving key insights."
},
"explain": {
"system": (
"You are an educator specializing in making complex topics simple. "
"Explain concepts for beginners using analogies, progressive "
"explanations, and clear examples. Avoid jargon unless necessary."
),
"description": "Explains complex topics in simple, beginner-friendly terms."
},
"creative": {
"system": (
"You are a creative writer. Generate imaginative, engaging, "
"and original ideas with vivid language. Think outside the box "
"and surprise the user with creative perspectives."
),
"description": "Creative writing and ideation with imaginative perspectives."
},
"general": {
"system": (
"You are a helpful, knowledgeable AI assistant. Provide clear, "
"accurate, and useful responses. Be concise but thorough."
),
"description": "General-purpose reasoning and assistance."
},
}
@mcp.tool(name="ask_specialized_claude")
def ask_specialized_claude(
prompt: str,
task_type: str = "general",
max_tokens: int = 1024
) -> str:
"""
Call a specialized Claude model for targeted reasoning tasks.
This is "meta-AI" - Claude in Cursor calling another Claude instance
with specialized system prompts for specific tasks. Each task_type
activates a different Claude persona optimized for that use case.
Args:
prompt: User input for Claude to process
task_type: Specialization mode, one of:
- code_review: Security and quality code analysis
- summarize: Concise text summarization
- explain: Educational explanations (ELI5 style)
- creative: Creative writing and ideation
- general: General-purpose assistance
max_tokens: Response length cap (default 1024, max 4096)
Returns:
A formatted string response from the specialized Claude instance
Examples:
>>> ask_specialized_claude("Explain quantum computing", "explain")
>>> ask_specialized_claude("Review this code: def foo()...", "code_review")
>>> ask_specialized_claude("Summarize this article...", "summarize")
Raises:
ValueError: If task_type is not recognized
Exception: If API call fails (network, auth, rate limit, etc.)
"""
# Validate task_type
if task_type not in CLAUDE_PROFILES:
valid = ", ".join(CLAUDE_PROFILES.keys())
raise ValueError(
f"Invalid task_type '{task_type}'. Must be one of: {valid}"
)
# Validate max_tokens
if not (1 <= max_tokens <= 4096):
raise ValueError(
f"max_tokens must be between 1 and 4096, got {max_tokens}"
)
# Get the profile for this task type
profile = CLAUDE_PROFILES[task_type]
# Get cached Anthropic client
try:
client = get_anthropic_client()
except EnvironmentError as e:
return f"❌ Configuration Error: {str(e)}"
try:
# Call Claude with specialized system prompt
response = client.messages.create(
model="claude-sonnet-4-20250514", # Latest Claude model
system=profile["system"],
max_tokens=max_tokens,
messages=[{"role": "user", "content": prompt}],
)
# Extract response text
text = response.content[0].text.strip() if response.content else "[No content returned]"
# Format with header showing which mode was used
header = f"🤖 Specialized Claude ({task_type.replace('_', ' ').title()} Mode)"
separator = "─" * len(header)
return f"{header}\n{separator}\n\n{text}"
except Exception as e:
# Provide helpful error messages
error_msg = str(e)
# Check for common error types
if "rate_limit" in error_msg.lower():
return (
f"❌ Rate Limit Error: Too many requests to Claude API.\n"
f"Please wait a moment and try again."
)
elif "authentication" in error_msg.lower() or "api_key" in error_msg.lower():
return (
f"❌ Authentication Error: Invalid ANTHROPIC_API_KEY.\n"
f"Please check your .env file and verify your API key."
)
elif "timeout" in error_msg.lower():
return (
f"❌ Timeout Error: Request to Claude API timed out.\n"
f"Please try again."
)
else:
return (
f"❌ Error invoking Claude ({task_type}): {error_msg}\n\n"
f"If this persists, check your API key and network connection."
)
# Self-registration confirmation
# When this module is imported, this tool is automatically registered
# with the shared MCP instance from core.mcp_instance
print("✅ Anthropic tools registered (ask_specialized_claude)")