Tracon.Anthropic 1.0.0-preview.6

Tracon.Anthropic

Anthropic (Claude) provider adapter for Tracon.

builder.AddTracon()
       .UseAnthropic(apiKey, o =>
       {
           o.DefaultModel = "claude-sonnet-5";
           o.DefaultMaxOutputTokens = 4096;
       });

Registered provider name: anthropic (AnthropicProviderNames.Anthropic).

Model = new ModelBinding
{
    Provider = AnthropicProviderNames.Anthropic,
    Model = "claude-sonnet-5",
    MaxOutputTokens = 1024,
}

Every produced IChatClient goes through the same pipeline as Tracon.OpenAI: UseFunctionInvocation() leaves the tool-call loop to the Microsoft Agent Framework, UseOpenTelemetry() produces spans under the Tracon source. Circuit-breaker and content-filter detection live at the ModelProviderRegistry level; this package gets both without extra code.

SDK used

The official Anthropic package (owner: Anthropic, MIT). The package carries its own AsIChatClient adapter, so message mapping, streaming, tool calling and usage counters are not written in Tracon. Its transitive dependencies are only Microsoft.Extensions.AI.Abstractions, System.Net.ServerSentEvents, System.Text.Json and System.IO.Pipelines.

Not Native AOT-compatible. The package's own code is analyzer-clean (IsAotCompatible=true, zero warnings), but the Anthropic SDK's IChatClient adapter initializes reflection-based JSON serialization: a Native AOT app that calls the provider fails on the first request with Reflection-based serialization has been disabled, and the publish emits thousands of trim and AOT warnings from the SDK. Measured 2026-10-17 with SDK 12.58.0.

max_tokens is required

In the Anthropic Messages API, max_tokens is a required field; it cannot be omitted the way it can in OpenAI. When ModelBinding.MaxOutputTokens is left empty, AnthropicProviderOptions.DefaultMaxOutputTokens (default 4096) is used.

A response that reaches the limit stops in the middle. A structured (Json or JsonSchema) response then fails the run as output_truncated; a text answer completes with a tracon.output.truncated run event. Thinking tokens count against the same limit, so set MaxOutputTokens explicitly for long answers and at high effort. When the catalog states the model's MaxOutputTokens, POST /api/agents/validate warns about a structured or high-effort agent left on the default.

A response the model stopped because the conversation filled its context window (model_context_window_exceeded) is a cut-off answer too. The error message and the event's cause (context_window) name the window: shorten the conversation or use a model with a larger window; a larger MaxOutputTokens does not help.

A turn the API paused (pause_turn, a long turn of its own server tools such as web search) is continued: the paused answer goes back as the API returned it, server tool blocks included, and the model goes on. Each call is billed and counted. A turn still paused after five continuations fails the call.

Provider-specific settings

Given per agent through the ModelBinding.ProviderSettings dictionary. A key absent from this list is not silently ignored — it fails compilation and the message lists the supported keys.

A wrong value is an error that names the key and the valid values. ttl and placement without anthropic.promptCaching = true, enabled without a budget, and adaptive with a budget are errors too.

var binding = new ModelBinding
{
    Provider = AnthropicProviderNames.Anthropic,
    Model = "your-current-claude-model",
    MaxOutputTokens = 64000,
    ProviderSettings = new Dictionary<string, JsonElement>(StringComparer.OrdinalIgnoreCase)
    {
        [AnthropicProviderNames.EffortSetting] = JsonSerializer.SerializeToElement("high"),
        [AnthropicProviderNames.ThinkingModeSetting] = JsonSerializer.SerializeToElement("adaptive"),
        [AnthropicProviderNames.PromptCachingSetting] = JsonSerializer.SerializeToElement(true),
        [AnthropicProviderNames.PromptCachingTtlSetting] = JsonSerializer.SerializeToElement("1h"),
        [AnthropicProviderNames.PromptCachingPlacementSetting] = JsonSerializer.SerializeToElement("system"),
    },
};

Known behavior differences

• Current models reject a thinking budget and sampling parameters with 400. Use anthropic.thinking.mode = adaptive with anthropic.effort, leave Temperature and TopP empty, and set SupportsSampling = false on the model's catalog entry so such a definition fails at compile time instead of at run time.
• ModelBinding.ReasoningEffort becomes adaptive thinking. It is sent as thinking: {"type":"adaptive"} plus output_config.effort (low, medium, high, xhigh), which every current Claude model accepts. A model that predates adaptive thinking (claude-haiku-4-5) rejects it with 400: select that model with anthropic.thinking.mode = enabled and anthropic.thinking.budgetTokens. A thinking provider setting wins over ReasoningEffort. Measured live on 2026-10-17 across the Claude 5 family and claude-haiku-4-5.
• Temperature must be 1 while budgeted thinking is on. When anthropic.thinking.budgetTokens is given, ModelBinding.Temperature must be either left empty or set to 1; any other value gets the request rejected with invalid_request_error.
• The thinking budget must be smaller than MaxOutputTokens.
• Prompt caching requires a threshold. Short prompts are not cached. The request placement treats the whole request as one cache unit, so a changed document or question writes a new entry. The system placement caches only up to the end of the instructions, which requests with different documents share. The effect is visible in the cached-input token counters of the run's usage.
• Provider settings do not reach a fallback link. A Fallbacks link gets the binding's portable fields only.
• The system message is a separate field. Anthropic carries the system field outside the message list; the SDK adapter does the conversion, Tracon has no work to do here.

Health check

Calls GET {endpoint}/models, incurs no cost. Authentication is done through the x-api-key header and the anthropic-version: 2023-06-01 header is required. The error detail is limited to the HTTP status code and a short reason; the API key or the endpoint address never leaks.

GET /tracon/api/models/health/anthropic

Model catalog

Tracon carries no built-in model list: model names change faster than a NuGet release. The catalog comes entirely from configuration and is not a validation list — a model name absent from it can still be used.

{
  "Tracon": {
    "Providers": {
      "Anthropic": {
        "ApiKey": "",
        "DefaultModel": "claude-sonnet-5",
        "DefaultMaxOutputTokens": 4096,
        "Models": [
          { "Name": "claude-sonnet-5", "DisplayName": "Claude Sonnet 5",
            "ContextWindowTokens": 200000, "MaxOutputTokens": 128000,
            "SupportsReasoning": true, "SupportsSampling": false,
            "InputCostPerMillionTokens": 3, "OutputCostPerMillionTokens": 15 }
        ]
      }
    }
  }
}

ApiKey is never written to this file — use dotnet user-secrets.

Links

• Guide: https://tracon.dev/guides/model-providers/
• Capability map: https://tracon.dev/capabilities/
• API reference: https://tracon.dev/api/

Licence: PolyForm Small Business 1.0.0 - free below 100 people and 1,000,000 USD (2019, inflation adjusted) revenue; a commercial licence applies above that. Terms ship in the package as LICENSE.md. Details: https://tracon.dev/reference/licensing/