Model Routing

AgentBridge standardizes the app-to-agent-framework layer. It does not try to become a second model gateway. In v0, model selection should stay in:

• AgentSpec.model, using LiteLLM-style or backend-supported model strings.
• Backend-specific extension config for native model settings.
• Environment variables or SDK/client settings for API keys, base URLs, and provider credentials.

Common Routes

Route	Example AgentSpec.model	Extra Settings
Offline CI	agentbridge/offline	No credentials; adapter-owned test model when supported.
OpenAI API	openai/gpt-5	OPENAI_API_KEY or backend-native auth.
Anthropic API	anthropic/claude-sonnet	ANTHROPIC_API_KEY or LiteLLM/provider settings.
Google API	google/gemini	Google credentials or ADK-native model settings.
Local Ollama-style model	ollama/llama3.1	Local model server running; backend must support that route.
OpenRouter	openrouter/openai/gpt-4o-mini	OPENROUTER_API_KEY, or OpenAI-compatible base_url in backend-native settings.
NVIDIA NIM	openai/deepseek-ai/deepseek-v4-flash-0731	NVIDIA_NIM_API_KEY, NVIDIA_NIM_API_BASE or NVIDIA_NIM_BASE_URL, optional NVIDIA_MODEL.
Custom gateway	openai/internal-agent-model	Internal OpenAI-compatible base URL and API key.

Backend Responsibilities

Adapters should:

• Pass AgentSpec.model through without rewriting provider semantics unless the backend requires it.
• Document which model string formats the backend has actually tested.
• Preserve native model settings in extension summaries when supplied.
• Keep credentialed provider tests optional and gated behind environment variables.
• Use agentbridge/offline for default CI whenever possible.

Adapters should not:

• Hide provider-specific model behavior behind fake portability.
• Require paid provider credentials for the default test suite.
• Claim that OpenAI-compatible endpoints work unless a smoke test or documented manual run exists.

For the LangChain plugin, an OpenAI-compatible native model can be configured without changing the framework-neutral AgentSpec shape:

agent = LangChainExtension.with_config(
    agent,
    model_provider="openai",
    model_options={
        "base_url": "https://integrate.api.nvidia.com/v1",
        "api_key": os.environ["NVIDIA_NIM_API_KEY"],
    },
)

Install agentbridge-langchain[openai] for this path. The adapter constructs the native ChatOpenAI object, passes it to LangChain, and redacts secrets from RunResult.metadata.

The same configuration shape supports provider-specific LangChain integrations without installing all providers into AgentBridge:

agent = LangChainExtension.with_config(
    agent,
    model_provider="anthropic",
    model_options={"temperature": 0.2},
)

Lazy factories cover Anthropic, Google GenAI/Vertex AI, Mistral, Groq, Cohere, Bedrock, Fireworks, Hugging Face, xAI, Azure OpenAI, OpenAI-compatible gateways, and local Ollama. Install the native langchain-* package for the provider and keep credentials in its normal environment/configuration. For provider behavior not represented by the factory, pass a constructed native chat model through model=.

Executable Provider Smoke Matrix

The repository includes a double-gated smoke matrix for NVIDIA NIM, OpenRouter, and local Ollama:

AGENTBRIDGE_RUN_CREDENTIAL_SMOKE=1 python examples/live_model_smoke.py

The callable form is run_provider_smoke_matrix_from_env(). It never runs by default. Configure NVIDIA_NIM_API_KEY plus NVIDIA_NIM_API_BASE or NVIDIA_NIM_BASE_URL for NIM, OPENROUTER_API_KEY for OpenRouter, or a reachable Ollama server. Optional overrides are NVIDIA_MODEL, OPENROUTER_BASE_URL, OPENROUTER_MODEL, OLLAMA_BASE_URL, and OLLAMA_MODEL. Each result records normalized content, usage, and provider errors without exposing API keys.

Observability providers use a separate executable lane:

AGENTBRIDGE_RUN_CREDENTIAL_SMOKE=1 \
  LANGSMITH_API_KEY=... \
  LANGFUSE_PUBLIC_KEY=... \
  LANGFUSE_SECRET_KEY=... \
  python examples/observability_smoke.py

This authenticates to LangSmith's /info endpoint and runs an offline LangChain invocation through the Langfuse callback path. It does not claim hosted evaluation, prompt-management, or deployment parity; those remain separate LangSmith API contract surfaces.

OpenAI-Compatible Providers

Many providers expose OpenAI-compatible chat/completions APIs. AgentBridge should represent these as a model string plus backend-native endpoint settings.

Example report shape:

{
  "model": "openai/deepseek-ai/deepseek-v4-flash-0731",
  "provider": "nvidia_nim",
  "base_url_env": "NVIDIA_NIM_API_BASE",
  "alternate_base_url_env": "NVIDIA_NIM_BASE_URL",
  "api_key_env": "NVIDIA_NIM_API_KEY",
  "model_env": "NVIDIA_MODEL",
  "credentialed": true,
  "default_ci": false
}

This keeps the model routing story explicit without adding another abstraction layer.

Provider smoke outputs can be compared without exposing credentials:

from examples.provider_comparison_report import build_provider_comparison_report

report = build_provider_comparison_report(route_results)

The report is result-driven and offline, making it suitable for CI artifacts and migration comparisons across NVIDIA NIM, OpenRouter, Ollama, and other compatible routes.

Scenario Report Requirement

Deep scenario reports should include a model_routes section showing:

• The offline route used by CI.
• Hosted provider routes that are documented or tested.
• Local model routes.
• OpenAI-compatible endpoint routes such as OpenRouter, NVIDIA NIM, and internal gateways.
• Which routes are metadata-only, no-key runnable, or credential-gated smoke tests.