29 KiB
29 KiB
Adding a New Provider
This guide explains how to add support for a new AI model provider to the Zen MCP Server. Follow these steps to integrate providers like Anthropic, Cohere, or any API that provides AI model access.
Overview
The provider system in Zen MCP Server is designed to be extensible. Each provider:
- Inherits from a base class (
ModelProviderorOpenAICompatibleProvider) - Implements required methods for model interaction
- Is registered in the provider registry by the server
- Has its API key configured via environment variables
Implementation Paths
You have two options when implementing a new provider:
Option A: Native Provider (Full Implementation)
Inherit from ModelProvider when:
- Your API has unique features not compatible with OpenAI's format
- You need full control over the implementation
- You want to implement custom features like extended thinking
Option B: OpenAI-Compatible Provider (Simplified)
Inherit from OpenAICompatibleProvider when:
- Your API follows OpenAI's chat completion format
- You want to reuse existing implementation for most functionality
- You only need to define model capabilities and validation
⚠️ CRITICAL: If your provider has model aliases (shorthands), you MUST override generate_content() to resolve aliases before API calls. See implementation example below.
Step-by-Step Guide
1. Add Provider Type to Enum
First, add your provider to the ProviderType enum in providers/base.py:
class ProviderType(Enum):
"""Supported model provider types."""
GOOGLE = "google"
OPENAI = "openai"
OPENROUTER = "openrouter"
CUSTOM = "custom"
EXAMPLE = "example" # Add your provider here
2. Create the Provider Implementation
Option A: Native Provider Implementation
Create a new file in the providers/ directory (e.g., providers/example.py):
"""Example model provider implementation."""
import logging
from typing import Optional
from .base import (
ModelCapabilities,
ModelProvider,
ModelResponse,
ProviderType,
RangeTemperatureConstraint,
)
from utils.model_restrictions import get_restriction_service
logger = logging.getLogger(__name__)
class ExampleModelProvider(ModelProvider):
"""Example model provider implementation."""
SUPPORTED_MODELS = {
"example-large-v1": {
"context_window": 100_000,
"supports_extended_thinking": False,
},
"example-small-v1": {
"context_window": 50_000,
"supports_extended_thinking": False,
},
# Shorthands
"large": "example-large-v1",
"small": "example-small-v1",
}
def __init__(self, api_key: str, **kwargs):
super().__init__(api_key, **kwargs)
# Initialize your API client here
def get_capabilities(self, model_name: str) -> ModelCapabilities:
resolved_name = self._resolve_model_name(model_name)
if resolved_name not in self.SUPPORTED_MODELS:
raise ValueError(f"Unsupported model: {model_name}")
restriction_service = get_restriction_service()
if not restriction_service.is_allowed(ProviderType.EXAMPLE, resolved_name, model_name):
raise ValueError(f"Model '{model_name}' is not allowed by restriction policy.")
config = self.SUPPORTED_MODELS[resolved_name]
return ModelCapabilities(
provider=ProviderType.EXAMPLE,
model_name=resolved_name,
friendly_name="Example",
context_window=config["context_window"],
supports_extended_thinking=config["supports_extended_thinking"],
supports_system_prompts=True,
supports_streaming=True,
supports_function_calling=True,
temperature_constraint=RangeTemperatureConstraint(0.0, 2.0, 0.7),
)
def generate_content(
self,
prompt: str,
model_name: str,
system_prompt: Optional[str] = None,
temperature: float = 0.7,
max_output_tokens: Optional[int] = None,
**kwargs,
) -> ModelResponse:
resolved_name = self._resolve_model_name(model_name)
self.validate_parameters(resolved_name, temperature)
# Call your API here
# response = your_api_call(...)
return ModelResponse(
content="", # From API response
usage={
"input_tokens": 0,
"output_tokens": 0,
"total_tokens": 0,
},
model_name=resolved_name,
friendly_name="Example",
provider=ProviderType.EXAMPLE,
)
def count_tokens(self, text: str, model_name: str) -> int:
# Implement your tokenization or use estimation
return len(text) // 4
def get_provider_type(self) -> ProviderType:
return ProviderType.EXAMPLE
def validate_model_name(self, model_name: str) -> bool:
resolved_name = self._resolve_model_name(model_name)
if resolved_name not in self.SUPPORTED_MODELS or not isinstance(self.SUPPORTED_MODELS[resolved_name], dict):
return False
restriction_service = get_restriction_service()
if not restriction_service.is_allowed(ProviderType.EXAMPLE, resolved_name, model_name):
logger.debug(f"Example model '{model_name}' -> '{resolved_name}' blocked by restrictions")
return False
return True
def supports_thinking_mode(self, model_name: str) -> bool:
capabilities = self.get_capabilities(model_name)
return capabilities.supports_extended_thinking
def _resolve_model_name(self, model_name: str) -> str:
shorthand_value = self.SUPPORTED_MODELS.get(model_name)
if isinstance(shorthand_value, str):
return shorthand_value
return model_name
Option B: OpenAI-Compatible Provider Implementation
For providers with OpenAI-compatible APIs, the implementation is much simpler:
"""Example provider using OpenAI-compatible interface."""
import logging
from typing import Optional
from .base import (
ModelCapabilities,
ModelResponse,
ProviderType,
RangeTemperatureConstraint,
)
from .openai_compatible import OpenAICompatibleProvider
logger = logging.getLogger(__name__)
class ExampleProvider(OpenAICompatibleProvider):
"""Example provider using OpenAI-compatible API."""
FRIENDLY_NAME = "Example"
# Define supported models
SUPPORTED_MODELS = {
"example-model-large": {
"context_window": 128_000,
"supports_extended_thinking": False,
},
"example-model-small": {
"context_window": 32_000,
"supports_extended_thinking": False,
},
# Shorthands
"large": "example-model-large",
"small": "example-model-small",
}
def __init__(self, api_key: str, **kwargs):
"""Initialize provider with API key."""
# Set your API base URL
kwargs.setdefault("base_url", "https://api.example.com/v1")
super().__init__(api_key, **kwargs)
def get_capabilities(self, model_name: str) -> ModelCapabilities:
"""Get capabilities for a specific model."""
resolved_name = self._resolve_model_name(model_name)
if resolved_name not in self.SUPPORTED_MODELS:
raise ValueError(f"Unsupported model: {model_name}")
# Check restrictions
from utils.model_restrictions import get_restriction_service
restriction_service = get_restriction_service()
if not restriction_service.is_allowed(ProviderType.EXAMPLE, resolved_name, model_name):
raise ValueError(f"Model '{model_name}' is not allowed by restriction policy.")
config = self.SUPPORTED_MODELS[resolved_name]
return ModelCapabilities(
provider=ProviderType.EXAMPLE,
model_name=resolved_name,
friendly_name=self.FRIENDLY_NAME,
context_window=config["context_window"],
supports_extended_thinking=config["supports_extended_thinking"],
supports_system_prompts=True,
supports_streaming=True,
supports_function_calling=True,
temperature_constraint=RangeTemperatureConstraint(0.0, 1.0, 0.7),
)
def get_provider_type(self) -> ProviderType:
"""Get the provider type."""
return ProviderType.EXAMPLE
def validate_model_name(self, model_name: str) -> bool:
"""Validate if the model name is supported."""
resolved_name = self._resolve_model_name(model_name)
if resolved_name not in self.SUPPORTED_MODELS or not isinstance(self.SUPPORTED_MODELS[resolved_name], dict):
return False
# Check restrictions
from utils.model_restrictions import get_restriction_service
restriction_service = get_restriction_service()
if not restriction_service.is_allowed(ProviderType.EXAMPLE, resolved_name, model_name):
return False
return True
def _resolve_model_name(self, model_name: str) -> str:
"""Resolve model shorthand to full name."""
shorthand_value = self.SUPPORTED_MODELS.get(model_name)
if isinstance(shorthand_value, str):
return shorthand_value
return model_name
def generate_content(
self,
prompt: str,
model_name: str,
system_prompt: Optional[str] = None,
temperature: float = 0.7,
max_output_tokens: Optional[int] = None,
**kwargs,
) -> ModelResponse:
"""Generate content using API with proper model name resolution."""
# CRITICAL: Resolve model alias before making API call
# This ensures aliases like "large" get sent as "example-model-large" to the API
resolved_model_name = self._resolve_model_name(model_name)
# Call parent implementation with resolved model name
return super().generate_content(
prompt=prompt,
model_name=resolved_model_name,
system_prompt=system_prompt,
temperature=temperature,
max_output_tokens=max_output_tokens,
**kwargs,
)
# Note: count_tokens is inherited from OpenAICompatibleProvider
3. Update Registry Configuration
3.1. Add Environment Variable Mapping
Update providers/registry.py to map your provider's API key:
@classmethod
def _get_api_key_for_provider(cls, provider_type: ProviderType) -> Optional[str]:
"""Get API key for a provider from environment variables."""
key_mapping = {
ProviderType.GOOGLE: "GEMINI_API_KEY",
ProviderType.OPENAI: "OPENAI_API_KEY",
ProviderType.OPENROUTER: "OPENROUTER_API_KEY",
ProviderType.CUSTOM: "CUSTOM_API_KEY",
ProviderType.EXAMPLE: "EXAMPLE_API_KEY", # Add this line
}
# ... rest of the method
4. Configure Docker Environment Variables
CRITICAL: You must add your provider's environment variables to docker-compose.yml for them to be available in the Docker container.
Add your API key and restriction variables to the environment section:
services:
zen-mcp:
# ... other configuration ...
environment:
- GEMINI_API_KEY=${GEMINI_API_KEY:-}
- OPENAI_API_KEY=${OPENAI_API_KEY:-}
- EXAMPLE_API_KEY=${EXAMPLE_API_KEY:-} # Add this line
# OpenRouter support
- OPENROUTER_API_KEY=${OPENROUTER_API_KEY:-}
# ... other variables ...
# Model usage restrictions
- OPENAI_ALLOWED_MODELS=${OPENAI_ALLOWED_MODELS:-}
- GOOGLE_ALLOWED_MODELS=${GOOGLE_ALLOWED_MODELS:-}
- EXAMPLE_ALLOWED_MODELS=${EXAMPLE_ALLOWED_MODELS:-} # Add this line
⚠️ Without this step, the Docker container won't have access to your environment variables, and your provider won't be registered even if the API key is set in your .env file.
5. Register Provider in server.py
The configure_providers() function in server.py handles provider registration. You need to:
Note: The provider priority is hardcoded in registry.py. If you're adding a new native provider (like Example), you'll need to update the PROVIDER_PRIORITY_ORDER in get_provider_for_model():
# In providers/registry.py
PROVIDER_PRIORITY_ORDER = [
ProviderType.GOOGLE, # Direct Gemini access
ProviderType.OPENAI, # Direct OpenAI access
ProviderType.EXAMPLE, # Add your native provider here
ProviderType.CUSTOM, # Local/self-hosted models
ProviderType.OPENROUTER, # Catch-all (must stay last)
]
Native providers should be placed BEFORE CUSTOM and OPENROUTER to ensure they get priority for their models.
- Import your provider class at the top of
server.py:
from providers.example import ExampleModelProvider
- Add API key checking in the
configure_providers()function:
def configure_providers():
"""Configure and validate AI providers based on available API keys."""
# ... existing code ...
# Check for Example API key
example_key = os.getenv("EXAMPLE_API_KEY")
if example_key and example_key != "your_example_api_key_here":
valid_providers.append("Example")
has_native_apis = True
logger.info("Example API key found - Example models available")
- Register the provider in the appropriate section:
# Register providers in priority order:
# 1. Native APIs first (most direct and efficient)
if has_native_apis:
if gemini_key and gemini_key != "your_gemini_api_key_here":
ModelProviderRegistry.register_provider(ProviderType.GOOGLE, GeminiModelProvider)
if openai_key and openai_key != "your_openai_api_key_here":
ModelProviderRegistry.register_provider(ProviderType.OPENAI, OpenAIModelProvider)
if example_key and example_key != "your_example_api_key_here":
ModelProviderRegistry.register_provider(ProviderType.EXAMPLE, ExampleModelProvider)
- Update error message to include your provider:
if not valid_providers:
raise ValueError(
"At least one API configuration is required. Please set either:\n"
"- GEMINI_API_KEY for Gemini models\n"
"- OPENAI_API_KEY for OpenAI o3 model\n"
"- EXAMPLE_API_KEY for Example models\n" # Add this
"- OPENROUTER_API_KEY for OpenRouter (multiple models)\n"
"- CUSTOM_API_URL for local models (Ollama, vLLM, etc.)"
)
6. Add Model Capabilities for Auto Mode
Update config.py to add your models to MODEL_CAPABILITIES_DESC:
MODEL_CAPABILITIES_DESC = {
# ... existing models ...
# Example models - Available when EXAMPLE_API_KEY is configured
"large": "Example Large (100K context) - High capacity model for complex tasks",
"small": "Example Small (50K context) - Fast model for simple tasks",
# Full model names
"example-large-v1": "Example Large (100K context) - High capacity model",
"example-small-v1": "Example Small (50K context) - Fast lightweight model",
}
7. Update Documentation
7.1. Update README.md
Add your provider to the quickstart section:
### 1. Get API Keys (at least one required)
**Option B: Native APIs**
- **Gemini**: Visit [Google AI Studio](https://makersuite.google.com/app/apikey)
- **OpenAI**: Visit [OpenAI Platform](https://platform.openai.com/api-keys)
- **Example**: Visit [Example API Console](https://example.com/api-keys) # Add this
Also update the .env file example:
# Edit .env to add your API keys
# GEMINI_API_KEY=your-gemini-api-key-here
# OPENAI_API_KEY=your-openai-api-key-here
# EXAMPLE_API_KEY=your-example-api-key-here # Add this
8. Write Tests
8.1. Unit Tests
Create tests/test_example_provider.py:
"""Tests for Example provider implementation."""
import os
from unittest.mock import patch
import pytest
from providers.example import ExampleModelProvider
from providers.base import ProviderType
class TestExampleProvider:
"""Test Example provider functionality."""
@patch.dict(os.environ, {"EXAMPLE_API_KEY": "test-key"})
def test_initialization(self):
"""Test provider initialization."""
provider = ExampleModelProvider("test-key")
assert provider.api_key == "test-key"
assert provider.get_provider_type() == ProviderType.EXAMPLE
def test_model_validation(self):
"""Test model name validation."""
provider = ExampleModelProvider("test-key")
# Test valid models
assert provider.validate_model_name("large") is True
assert provider.validate_model_name("example-large-v1") is True
# Test invalid model
assert provider.validate_model_name("invalid-model") is False
def test_resolve_model_name(self):
"""Test model name resolution."""
provider = ExampleModelProvider("test-key")
# Test shorthand resolution
assert provider._resolve_model_name("large") == "example-large-v1"
assert provider._resolve_model_name("small") == "example-small-v1"
# Test full name passthrough
assert provider._resolve_model_name("example-large-v1") == "example-large-v1"
def test_get_capabilities(self):
"""Test getting model capabilities."""
provider = ExampleModelProvider("test-key")
capabilities = provider.get_capabilities("large")
assert capabilities.model_name == "example-large-v1"
assert capabilities.friendly_name == "Example"
assert capabilities.context_window == 100_000
assert capabilities.provider == ProviderType.EXAMPLE
# Test temperature range
assert capabilities.temperature_constraint.min_temp == 0.0
assert capabilities.temperature_constraint.max_temp == 2.0
8.2. Simulator Tests (Real-World Validation)
Create a simulator test to validate that your provider works correctly in real-world scenarios. Create simulator_tests/test_example_models.py:
"""
Example Provider Model Tests
Tests that verify Example provider functionality including:
- Model alias resolution
- API integration
- Conversation continuity
- Error handling
"""
from .base_test import BaseSimulatorTest
class TestExampleModels(BaseSimulatorTest):
"""Test Example provider functionality"""
@property
def test_name(self) -> str:
return "example_models"
@property
def test_description(self) -> str:
return "Example provider model functionality and integration"
def run_test(self) -> bool:
"""Test Example provider models"""
try:
self.logger.info("Test: Example provider functionality")
# Check if Example API key is configured
check_result = self.check_env_var("EXAMPLE_API_KEY")
if not check_result:
self.logger.info(" ⚠️ Example API key not configured - skipping test")
return True # Skip, not fail
# Test 1: Shorthand alias mapping
self.logger.info(" 1: Testing 'large' alias mapping")
response1, continuation_id = self.call_mcp_tool(
"chat",
{
"prompt": "Say 'Hello from Example Large model!' and nothing else.",
"model": "large", # Should map to example-large-v1
"temperature": 0.1,
}
)
if not response1:
self.logger.error(" ❌ Large alias test failed")
return False
self.logger.info(" ✅ Large alias call completed")
# Test 2: Direct model name
self.logger.info(" 2: Testing direct model name (example-small-v1)")
response2, _ = self.call_mcp_tool(
"chat",
{
"prompt": "Say 'Hello from Example Small model!' and nothing else.",
"model": "example-small-v1",
"temperature": 0.1,
}
)
if not response2:
self.logger.error(" ❌ Direct model name test failed")
return False
self.logger.info(" ✅ Direct model name call completed")
# Test 3: Conversation continuity
self.logger.info(" 3: Testing conversation continuity")
response3, new_continuation_id = self.call_mcp_tool(
"chat",
{
"prompt": "Remember this number: 99. What number did I just tell you?",
"model": "large",
"temperature": 0.1,
}
)
if not response3 or not new_continuation_id:
self.logger.error(" ❌ Failed to start conversation")
return False
# Continue conversation
response4, _ = self.call_mcp_tool(
"chat",
{
"prompt": "What was the number I told you earlier?",
"model": "large",
"continuation_id": new_continuation_id,
"temperature": 0.1,
}
)
if not response4:
self.logger.error(" ❌ Failed to continue conversation")
return False
if "99" in response4:
self.logger.info(" ✅ Conversation continuity working")
else:
self.logger.warning(" ⚠️ Model may not have remembered the number")
# Test 4: Check logs for proper provider usage
self.logger.info(" 4: Validating Example provider usage in logs")
logs = self.get_recent_server_logs()
# Look for evidence of Example provider usage
example_logs = [line for line in logs.split("\n") if "example" in line.lower()]
model_resolution_logs = [
line for line in logs.split("\n")
if "Resolved model" in line and "example" in line.lower()
]
self.logger.info(f" Example-related logs: {len(example_logs)}")
self.logger.info(f" Model resolution logs: {len(model_resolution_logs)}")
# Success criteria
api_used = len(example_logs) > 0
models_resolved = len(model_resolution_logs) > 0
if api_used and models_resolved:
self.logger.info(" ✅ Example provider tests passed")
return True
else:
self.logger.error(" ❌ Example provider tests failed")
return False
except Exception as e:
self.logger.error(f"Example provider test failed: {e}")
return False
def main():
"""Run the Example provider tests"""
import sys
verbose = "--verbose" in sys.argv or "-v" in sys.argv
test = TestExampleModels(verbose=verbose)
success = test.run_test()
sys.exit(0 if success else 1)
if __name__ == "__main__":
main()
The simulator test is crucial because it:
- Validates your provider works in the actual Docker environment
- Tests real API integration, not just mocked behavior
- Verifies model name resolution works correctly
- Checks conversation continuity across requests
- Examines server logs to ensure proper provider selection
See simulator_tests/test_openrouter_models.py for a complete real-world example.
Model Name Mapping and Provider Priority
How Model Name Resolution Works
When a user requests a model (e.g., "pro", "o3", "example-large-v1"), the system:
-
Checks providers in priority order (defined in
registry.py):PROVIDER_PRIORITY_ORDER = [ ProviderType.GOOGLE, # Native Gemini API ProviderType.OPENAI, # Native OpenAI API ProviderType.CUSTOM, # Local/self-hosted ProviderType.OPENROUTER, # Catch-all for everything else ] -
For each provider, calls
validate_model_name():- Native providers (Gemini, OpenAI) return
trueonly for their specific models - OpenRouter returns
truefor ANY model (it's the catch-all) - First provider that validates the model handles the request
- Native providers (Gemini, OpenAI) return
Example: Model "gemini-2.5-pro"
- Gemini provider checks: YES, it's in my SUPPORTED_MODELS → Gemini handles it
- OpenAI skips (Gemini already handled it)
- OpenRouter never sees it
Example: Model "claude-3-opus"
- Gemini provider checks: NO, not my model → skip
- OpenAI provider checks: NO, not my model → skip
- Custom provider checks: NO, not configured → skip
- OpenRouter provider checks: YES, I accept all models → OpenRouter handles it
Implementing Model Name Validation
Your provider's validate_model_name() should:
def validate_model_name(self, model_name: str) -> bool:
resolved_name = self._resolve_model_name(model_name)
# Only accept models you explicitly support
if resolved_name not in self.SUPPORTED_MODELS or not isinstance(self.SUPPORTED_MODELS[resolved_name], dict):
return False
# Check restrictions
restriction_service = get_restriction_service()
if not restriction_service.is_allowed(ProviderType.EXAMPLE, resolved_name, model_name):
logger.debug(f"Example model '{model_name}' -> '{resolved_name}' blocked by restrictions")
return False
return True
Important: Native providers should ONLY return true for models they explicitly support. This ensures they get priority over proxy providers like OpenRouter.
Model Shorthands
Each provider can define shorthands in their SUPPORTED_MODELS:
SUPPORTED_MODELS = {
"example-large-v1": { ... }, # Full model name
"large": "example-large-v1", # Shorthand mapping
}
The _resolve_model_name() method handles this mapping automatically.
Critical Implementation Requirements
Alias Resolution for OpenAI-Compatible Providers
If you inherit from OpenAICompatibleProvider and define model aliases, you MUST override generate_content() to resolve aliases before API calls. This is because:
- The base
OpenAICompatibleProvider.generate_content()sends the original model name directly to the API - Your API expects the full model name, not the alias
- Without resolution, requests like
model="large"will fail with 404/400 errors
Examples of providers that need this:
- XAI provider:
"grok"→"grok-3" - OpenAI provider:
"mini"→"o4-mini" - Custom provider:
"fast"→"llama-3.1-8b-instruct"
Example implementation pattern:
def generate_content(self, prompt: str, model_name: str, **kwargs) -> ModelResponse:
# CRITICAL: Resolve alias before API call
resolved_model_name = self._resolve_model_name(model_name)
# Pass resolved name to parent
return super().generate_content(prompt=prompt, model_name=resolved_model_name, **kwargs)
Providers that DON'T need this:
- Gemini provider (has its own generate_content implementation)
- OpenRouter provider (already implements this pattern)
- Providers without aliases
Best Practices
- Always validate model names against supported models and restrictions
- Be specific in validation - only accept models you actually support
- Handle API errors gracefully with proper error messages
- Include retry logic for transient errors (see
gemini.pyfor example) - Log important events for debugging (initialization, model resolution, errors)
- Support model shorthands for better user experience
- Document supported models clearly in your provider class
- Test thoroughly including error cases and edge conditions
Checklist
Before submitting your PR:
- Provider type added to
ProviderTypeenum inproviders/base.py - Provider implementation complete with all required methods
- API key mapping added to
_get_api_key_for_provider()inproviders/registry.py - Provider added to
PROVIDER_PRIORITY_ORDERinregistry.py(if native provider) - Environment variables added to
docker-compose.yml(API key and restrictions) - Provider imported and registered in
server.py'sconfigure_providers() - API key checking added to
configure_providers()function - Error message updated to include new provider
- Model capabilities added to
config.pyfor auto mode - Documentation updated (README.md)
- Unit tests written and passing (
tests/test_<provider>.py) - Simulator tests written and passing (
simulator_tests/test_<provider>_models.py) - Integration tested with actual API calls
- Code follows project style (run linting)
- PR follows the template requirements
Need Help?
- Look at existing providers (
gemini.py,openai.py) for examples - Check the base classes for method signatures and requirements
- Run tests frequently during development
- Ask questions in GitHub issues if stuck