Choose a provider¶

AgentMuru 0.3 ships three adapters built on the official provider SDKs. They translate provider messages and streams into the same runtime events, so tool governance and persistence remain provider-neutral.

Provider	Extra	Default model	Guide
OpenAI Responses	`agentmuru[openai]`	`gpt-5.6-terra`	Configure OpenAI
Anthropic Messages	`agentmuru[anthropic]`	`claude-sonnet-5`	Configure Anthropic
Google Gen AI	`agentmuru[google]`	`gemini-3.5-flash`	Configure Google
Deterministic local fake	Core package	`fake`	Agents and models

Stable runtime contract¶

All adapters:

stream TextDelta events;
emit ToolCall only with complete object arguments;
report normalized input and output token usage;
map provider failures to stable, secret-safe ModelFailed codes;
propagate cancellation and close active streams;
preserve prior assistant tool calls and tool results on the next turn.

Install one SDK¶

python -m pip install "agentmuru[openai]>=0.3,<0.4"

Use agentmuru[providers] only when one environment must exercise every adapter. Keeping one provider extra in an application reduces its dependency and update surface.

Switch providers¶

Change the import and model constructor. Do not change the agent, tools, runtime, or stores.

from agentmuru.integrations.openai import OpenAIModel

agent = Agent(
    name="assistant",
    instructions="Answer clearly.",
    model=OpenAIModel(),
)

Each model constructor accepts an explicit model ID. Tests may inject a provider client to avoid credentials and network calls.

Errors¶

Authentication, permission, rate limit, timeout, invalid request, unavailable service, invalid tool arguments, and unclassified provider failures map to documented AgentMuru codes. Provider response bodies and API keys never become public failure messages.

See the provider contract reference for normalized settings and error codes.