> ## Documentation Index
> Fetch the complete documentation index at: https://docs.meshagent.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Anthropic Messages Adapter

The Anthropic Messages adapter is an `LLMAdapter` implementation for Anthropic's Messages API. It enables MeshAgent agents to use Claude models, handle tool calls, and stream responses while preserving MeshAgent toolkit behavior.

## Key features

* **Model defaults:** Reads the model name from the constructor (`model=`) or the `ANTHROPIC_MODEL` environment variable. Override per call by passing `model` to `create_response()`.
* **Max token defaults:** Reads `max_tokens` from the constructor or `ANTHROPIC_MAX_TOKENS`.
* **Session context defaults:** Creates `AgentSessionContext(system_role=None)` so system/developer prompts come from the caller.
* **Tool calling:** Converts the supplied toolkits into Anthropic `tools` definitions and executes `tool_use` requests.
* **Tool result formatting:** Uses `AnthropicMessagesToolResponseAdapter` to return `tool_result` blocks back to the model.
* **Streaming support:** Uses Anthropic's streaming API and emits events to `event_handler`.
* **Structured output (best-effort):** When `output_schema` is provided, the adapter prompts for JSON and validates it (with retries).
* **MCP connector support:** MCP toolkits inject `mcp_servers`, MCP toolset entries in `tools`, and required `betas` flags into the request.

## Constructor parameters

```python Python theme={null}
AnthropicMessagesAdapter(
    model: str = "claude-3-5-sonnet-latest",
    max_tokens: int = 1024,
    client: Optional[Any] = None,
    message_options: Optional[dict] = None,
    provider: str = "anthropic",
    log_requests: bool = False,
    context_management: Literal["auto", "none"] = "auto",
    compaction_threshold: int = 150000,
    compaction_pause_after: bool = False,
    compaction_instructions: Optional[str] = None,
    base_url: Optional[str] = None,
)
```

* `model` - default model name; can be overridden per call.
* `max_tokens` - cap on output tokens per response.
* `client` - reuse an existing Anthropic client; otherwise the adapter builds one via `meshagent.anthropic.proxy.get_client`.
* `base_url` - override the provider base URL used when the adapter creates its own client. Defaults to `ANTHROPIC_BASE_URL` when omitted.
* `message_options` - extra parameters passed to `messages.create` (for example `temperature`, `top_p`, `tools`, `betas`).
* `provider` - label emitted in telemetry and logs.
* `log_requests` - when true, uses a logging HTTP client for debugging.
* `context_management` - `auto` enables Anthropic beta context management compaction edits only for Claude model versions newer than `4.5` (for example `claude-sonnet-4-6`); `none` disables automatic compaction configuration.
* `compaction_threshold` - input token threshold for triggering `compact_20260112` when `context_management="auto"` (minimum `50000`).
* `compaction_pause_after` - when true, compaction can return a `compaction` block and pause before continuing.
* `compaction_instructions` - optional summarization instructions passed to compaction.

When `context_management="auto"` is enabled, the adapter uses Anthropic's `compact-2026-01-12` beta on `client.beta.messages.*` and injects a `compact_20260112` edit into the request's `context_management`.

## MCP connector support

Anthropic's official MCP connector can be enabled by adding an MCP toolkit to your toolkits list. The adapter applies MCP middleware to the request, which injects `mcp_servers` and an MCP toolset into the top-level `tools` array (and ensures the required `betas` flag is present).

## Handling a turn

When `create_response()` is called it:

1. **Bundles tools** - Converts toolkits into Anthropic tool definitions.
2. **Builds the request** - Converts the chat context into Anthropic blocks and constructs the request payload.
3. **Calls the model** - Sends the request (streaming if `event_handler` is provided).
4. **Handles tool calls** - Executes requested tools and formats `tool_result` blocks.
5. **Loops** - Continues until the model returns a final response.
6. **Returns result** - Returns text or validated JSON if `output_schema` was supplied.

## Related Topics

* [Adapters Overview](./index): Understand LLMAdapters and ToolResponseAdapters
* [MeshAgent LLM Proxy](../routing/llm_proxy): How requests are routed and metered inside rooms
* [OpenAI Responses Adapter](./openai_responses_adapter): A detailed reference implementation
* [Anthropic Tool Response Adapter](./anthropic_tool_response_adapter): How tool outputs are rendered back into the chat transcript
* [Process Agents Overview](../process/overview): Shows where adapters are used in the recommended runtime