Zonit.Extensions.Ai.OpenAi 10.0.8

<div align="center">

Zonit.Extensions.Ai

One typed, AOT-ready .NET API for every major AI provider: text, images, audio, embeddings, agents and tools.

</div>

---

Zonit.Extensions.Ai talks to OpenAI, Anthropic, Google, xAI and a dozen other providers through a single interface, IAiProvider. You pick a model as a strongly-typed class, pass a typed prompt, and get back a typed result with usage and cost already calculated. The same GenerateAsync call resolves to text, image, audio or embedding by the model you hand it; agents and tool-driven chat use the fluent ai.Agent(...) / ai.Chat(...) builder. No separate clients, no stringly-typed configuration, no provider-specific SDKs in your domain code.

IAiProvider ai = /* injected */;

// Plain text
var answer = await ai.GenerateAsync(new GPT5(), "Summarise the CAP theorem in one sentence.");

// Typed, structured output (JSON Schema generated for you)
var review = await ai.GenerateAsync(new Sonnet46(), new CodeReviewPrompt { Diff = diff });
Console.WriteLine(review.Value.Severity);     // strongly typed
Console.WriteLine(review.MetaData.TotalCost); // cost, already computed

📖 Guides: Instruction/. Task-focused docs that are also compiled into your AI assistant on install (see AI-assistant ready).

Everything in one library

One API covers every provider, every modality, and the agentic stack.

16 providers	OpenAI · Anthropic · Google · xAI · DeepSeek · Mistral · Groq · Together · Fireworks · Cohere · Perplexity · Alibaba · Baidu · Zhipu · Moonshot · 01.AI
Every modality	text · structured output (JSON Schema) · images · embeddings · audio transcription · video · streaming
Agentic	typed tools · external MCP servers · parallel tool calls · streaming events · full audit trail
Production	token usage and computed cost (including nested) · retry, timeout, circuit-breaker · reasoning and fast modes · AOT- and trim-ready (net10.0)
AI-assistant ready	ships instructions for GitHub Copilot, Claude Code and Cursor; your coding agent learns the library on install (details)

Table of contents

• Why this library
• AI-assistant ready
• Installation
• Quick start
• Core concepts
• The unified API
• Working with modalities
• Reasoning and fast mode
• Prompt caching
• Multi-turn chat
• Agents and tools
• Ready-made prompts
• Cost tracking and estimation
• Configuration
• Resilience
• AOT and trimming
• Provider packages
• Architecture
• Requirements and license

---

Why this library

• One API, every provider. Inject IAiProvider once. Switch a model class to switch providers; the call site does not change.
• Typed prompts, typed answers. Define a PromptBase<TResponse>. The library generates the JSON Schema, sends it as a structured-output request, and deserializes the reply into TResponse.
• Scriban templating built in. Prompt properties are exposed to the template as snake_case, with loops, conditionals and interpolation.
• Agents with tools and MCP. Typed tools (ToolBase<TInput, TOutput>), external MCP servers over HTTP/SSE, parallel tool execution, streaming events, and a complete audit trail on ResultAgent<T>.
• Cost is a first-class value. Every result carries token usage and computed cost. Agent runs also roll up the cost of nested AI calls made inside your tools.
• Production resilience. Retry, timeout and circuit-breaker via Microsoft.Extensions.Http.Resilience, configurable globally or per provider.
• AOT- and trim-ready. Targets net10.0 and source-generates the structured-output path. See AOT and trimming for what is and is not AOT-safe.

---

AI-assistant ready

When you install the package, a build-time step compiles a complete description of the library into your repository, so GitHub Copilot, Claude Code and Cursor know how to use IAiProvider correctly without being prompted. That includes which provider NuGet to install for a given capability: ask for "generate images with GPT" and the assistant installs Zonit.Extensions.Ai.OpenAi and writes the code.

• Only for editors you use. Files are written only for editors detected in the repo (.claude/ or CLAUDE.md, .cursor/, .vscode/, or existing Copilot instructions). With no AI editor present, nothing is written.

• One source, compiled per editor. A single set of Markdown docs (the same files in Instruction/) is projected into each editor's native format:

  Editor	Output
  Cursor	.cursor/rules/zonit-ai*.mdc, scoped to *Prompt.cs and *Tool.cs via globs
  GitHub Copilot	.github/instructions/zonit-ai*.instructions.md (applyTo) and a block in copilot-instructions.md
  Claude Code	a Skill (.claude/skills/zonit-extensions-ai/) plus a pointer block in CLAUDE.md
  (all)	a guide tree under .zonit/extensions/ai/ and a map at .zonit/index.md

• Safe and idempotent. Blocks live between markers, so your own edits are never overwritten; nothing is rewritten unless it changed; the step is skipped on CI.

• Control: <ZonitAiInstructions>false</ZonitAiInstructions> disables it; <ZonitAiEditors>auto|all|none|cursor;claude</ZonitAiEditors> overrides detection.

---

Installation

Install the core package plus the provider package(s) you need:

# Core (required): DI, prompts, templating, agent runtime
dotnet add package Zonit.Extensions.Ai

# Add one or more providers
dotnet add package Zonit.Extensions.Ai.OpenAi
dotnet add package Zonit.Extensions.Ai.Anthropic
dotnet add package Zonit.Extensions.Ai.Google

The MCP client and agent runtime live in the core package; there is nothing extra to install for tools or agents. See Provider packages for the full list.

---

Quick start

1. Store your keys

// appsettings.json  (use User Secrets or environment variables in production)
{
  "Ai": {
    "OpenAi":    { "ApiKey": "sk-..." },
    "Anthropic": { "ApiKey": "sk-ant-..." }
  }
}

2. Register the providers

using Zonit.Extensions;

var builder = Host.CreateApplicationBuilder(args);

builder.Services.AddAiOpenAi();      // binds "Ai:OpenAi" automatically
builder.Services.AddAiAnthropic();   // binds "Ai:Anthropic"

var app = builder.Build();

AddAi{Provider}() registers the provider, binds its configuration section, and wires the core services and a resilient HttpClient. Registration is idempotent, so it is safe to call from as many modules or plugins as you like.

3. Inject IAiProvider and call

internal sealed class Summariser(IAiProvider ai)
{
    public async Task<string> RunAsync(string text, CancellationToken ct)
    {
        var result = await ai.GenerateAsync(new GPT5(), $"Summarise:\n{text}", ct);
        return result.Value;
    }
}

The whole loop: choose a model, pass a prompt, read result.Value.

---

Core concepts

Models are capabilities, not strings

A model is a small class that implements ILlm and one or more capability interfaces. The interface a model implements decides which GenerateAsync overload accepts it, so the compiler stops you from, for example, asking an embedding model to generate an image.

Capability interface	What it enables
ILlm	Base contract: name, context window, pricing, capabilities
IAgentLlm	Tool-calling and the agent loop (most chat models implement this)
IReasoningLlm	Configurable reasoning effort, summaries, verbosity
IImageLlm	Image generation
IEmbeddingLlm	Text embeddings
IAudioLlm	Audio transcription
IVideoLlm	Video generation
IFast	Opt-in fast inference tier (premium pricing)

Concrete model classes (GPT5, Sonnet46, GPTImage15, TextEmbedding3Large) ship inside each provider package and evolve with the providers. Discover them through IntelliSense or the package's Llm/ folder; each class carries its own context window, pricing and supported endpoints. This README lists no model tables, because they go stale every release.

Prompts are types

Inherit PromptBase<TResponse>. Public properties become both template variables and the data the model sees; TResponse becomes the structured-output contract.

using System.ComponentModel;
using Zonit.Extensions; // Culture value object

public sealed class SentimentPrompt : PromptBase<SentimentResponse>
{
    public required string Text { get; init; }

    public override string Prompt => """
        Classify the sentiment of the text below as positive, neutral or negative.

        {{ text }}
        """;
}

[Description("Sentiment classification.")]
public sealed class SentimentResponse
{
    [Description("positive, neutral or negative")]
    public required string Sentiment { get; init; }

    [Description("Confidence from 0.0 to 1.0.")]
    public double Confidence { get; init; }
}

[Description] attributes are forwarded into the JSON Schema, so the model receives your intent verbatim. For one-off calls without a class, use SimplePrompt<T>. For ready-made prompts, see Ready-made prompts.

Results carry usage and cost

Every call returns Result<T>:

public class Result<T>
{
    public required T        Value    { get; init; }  // the typed answer
    public required MetaData MetaData { get; init; }  // model, usage, cost, timing
}

MetaData exposes the model, provider, derived PromptName, Duration, RequestId, full TokenUsage, and shortcuts such as InputTokens, TotalTokens and TotalCost. Costs are Price value objects, computed from the model's own pricing.

---

The unified API

All generation goes through IAiProvider, on two surfaces: simple positional calls for plain in→out, and the fluent Agent / Chat builder for anything with tools, MCP, context or limits (safe by default — no tool reaches the model unless you add it). Simple calls, selected by the model's capability interface:

Call	Model type	Returns
GenerateAsync(llm, string) / (llm, IPrompt<T>)	ILlm	Result<string> / Result<T>
GenerateAsync(imageLlm, string) / (…, IPrompt<Asset>)	IImageLlm	Result<Asset>
GenerateAsync(videoLlm, string) / (…, IPrompt<Asset>)	IVideoLlm	Result<Asset>
GenerateAsync(embeddingLlm, string)	IEmbeddingLlm	Result<float[]>
GenerateAsync(audioLlm, Asset, language?)	IAudioLlm	Result<string>
ChatAsync(llm, prompt, history)	ILlm	Result<T> — multi-turn, no tools
StreamAsync(llm, string) · ChatStreamAsync(llm, prompt, history)	ILlm	IAsyncEnumerable<string> (tokens)
CalculateCost(...) / EstimateCost(...)	various	Price

For tools, MCP, scoped context or per-call limits, use the fluent builder — ai.Agent(agentLlm, prompt) or ai.Chat(llm, prompt, history), each terminated by .RunAsync() (awaited result) or .RunStreamAsync() (event stream). Full reference in Agents and tools. Every text entry point also has a plain-string overload.

---

Working with modalities

Text and structured output

// Free text
Result<string> reply = await ai.GenerateAsync(new GPT5(), "What is 2 + 2?");
Console.WriteLine(reply.Value); // "4"

// Structured. SentimentResponse is filled from the model's JSON.
Result<SentimentResponse> r = await ai.GenerateAsync(
    new Sonnet46(),
    new SentimentPrompt { Text = "I love this!" });

Console.WriteLine(r.Value.Sentiment);  // "positive"
Console.WriteLine(r.MetaData.PromptName);   // "Translate"

Streaming text

await foreach (var chunk in ai.StreamAsync(new GPT5(), "Tell me a short story."))
    Console.Write(chunk);

Files and vision

Asset (from Zonit.Extensions) detects the MIME type and converts implicitly from byte[] and Stream. Attach files to any prompt through Files:

public sealed class InvoicePrompt : PromptBase<InvoiceData>
{
    public override string Prompt => "Extract the invoice fields from the attached document.";
}

var bytes = await File.ReadAllBytesAsync("invoice.pdf");
var result = await ai.GenerateAsync(
    new GPT5(),
    new InvoicePrompt { Files = [new Asset(bytes, "invoice.pdf")] });

Image generation

Each image model declares its own QualityType and SizeType enums, mapped to the exact API values, so you cannot pass an unsupported combination.

var image = await ai.GenerateAsync(
    new GPTImage15
    {
        Quality = GPTImage15.QualityType.High,
        Size    = GPTImage15.SizeType.Landscape,
    },
    "A lighthouse at dusk, dramatic clouds, photorealistic");

await File.WriteAllBytesAsync("lighthouse.png", image.Value.Data);

Embeddings

Result<float[]> embedding = await ai.GenerateAsync(
    new TextEmbedding3Large(), "vectorise me");

float[] vector = embedding.Value;

Audio transcription

var audio = new Asset(await File.ReadAllBytesAsync("speech.mp3"), "speech.mp3");
var text  = await ai.GenerateAsync(new GPT4oTranscribe(), audio, language: "en");
Console.WriteLine(text.Value);

---

Reasoning and fast mode

Reasoning models expose effort, summary and verbosity controls through typed properties:

var result = await ai.GenerateAsync(
    new GPT52
    {
        Reason    = OpenAiReasoningBase.ReasonType.High,
        Verbosity = OpenAiReasoningBase.VerbosityType.Low,
    },
    new ProofPrompt { Statement = "root 2 is irrational" });

Some models offer a fast inference tier (IFast): the same weights at higher throughput and premium pricing. Cost calculation switches to the fast rate automatically when you select it.

var fast = await ai.GenerateAsync(
    new Opus48 { Speed = SpeedType.Fast },
    "Draft a release note for v10.");

---

Prompt caching

Anthropic models support prompt caching: the large, stable part of a request — the system prompt, the tool catalogue, the conversation so far — is cached server-side and replayed on later turns at a fraction of the input price. Enable it per model with the Cache property; it is off by default.

using Zonit.Extensions.Ai.Anthropic;   // the Cache enum

// Agents, chat loops and any repeated-prefix calls benefit most.
var result = await ai.Agent(
        new Opus48 { Cache = Cache.FiveMinutes },     // None (default) | FiveMinutes | OneHour
        new ResearchPrompt { Topic = "EU AI Act" })
    .AddTool<SearchTool>()
    .AddTool<SaveNoteTool>()
    .RunAsync();

TTL	Use it when
Cache.None	One-off calls with no shared prefix (default).
Cache.FiveMinutes	Agent and chat loops where turns land within a few minutes.
Cache.OneHour	Long-running sessions or chats with idle gaps over five minutes (beta TTL).

The economics: the first turn writes the prefix to the cache at a premium (1.25× input for FiveMinutes, 2× for OneHour), so it is slightly more expensive; every turn after that reads that prefix at ~10% of the input price. Caching is therefore net-positive from the second turn onward — exactly the shape of an agent run, where the system prompt and tool definitions repeat every turn. No per-call wiring is needed: once Cache is set, the library manages the cache breakpoints (tools, system, and the two most recent turns) automatically. Cached and cache-write tokens are reported separately on MetaData.Usage and priced at the model's cached rates, so TotalCost stays accurate. Caching is Anthropic-specific; other providers ignore the property.

---

Multi-turn chat

ChatAsync is the conversational counterpart to GenerateAsync. The conversation lives in a ChatMessage[] of User and Assistant records; the prompt supplies the system instruction. In chat mode prompt.Text is the system message, the inverse of single-shot GenerateAsync, where it is the user message.

var history = new ChatMessage[]
{
    new User("Why does my deployment hang?"),
    new Assistant("Often a missing source-gen override. What error do you see?"),
    new User("CS0534 on .NET 10."),
};

var result = await ai.ChatAsync(
    new Sonnet46(),
    new HelpdeskPrompt { Product = "Zonit.Ai" },   // system instruction
    history);

Console.WriteLine(result.Value);

For a tool-driven conversation, switch to the fluent builder — ai.Chat(llm, prompt, history).AddTool<T>()….RunAsync() — the turn then routes through the agent runner and the result's .Value is a ResultAgent<T> (see below). For token-by-token output without tools, use ChatStreamAsync. Native multi-turn message arrays are used for OpenAI, Anthropic, xAI (Grok) and Google (Gemini); other providers fall back to a flattened transcript.

---

Agents and tools

An agent drives the model in a loop: the model requests tool calls, the library runs them in parallel, feeds the results back, and repeats until the model returns a final structured answer, all behind one await. Prefer a single prompt for self-contained tasks and use an agent when the model needs to fetch data or take actions mid-task.

var result = await ai.Agent(new GPT5(), new ResearchPrompt { Topic = "EU AI Act" })  // any IAgentLlm
    .AddTool<GetWeatherTool>()                    // tools off by default — add what this call needs
    .AddTool<SaveNoteTool>()
    .AddMcp("github", "https://mcp.example.com/sse", token)
    .RunAsync();

Console.WriteLine($"Answer: {result.Value}");
Console.WriteLine($"Iterations: {result.Iterations}, cost: {result.Total.Cost}");

Authoring a tool

Tools are typed exactly like prompts. Inherit ToolBase<TInput, TOutput> and the library generates the input schema, deserializes the model's arguments, serializes your result, and traps exceptions.

public sealed class SaveNoteTool(INoteStore store)
    : ToolBase<SaveNoteTool.Input, SaveNoteTool.Output>
{
    public override string Name        => "save_note";
    public override string Description => "Persists a note and returns its id.";

    public override async Task<Output> ExecuteAsync(Input input, CancellationToken ct)
    {
        var id = await store.SaveAsync(input.Title, input.Body, ct);
        return new Output { Id = id };
    }

    public sealed class Input
    {
        [Description("Short title.")]      public required string Title { get; init; }
        [Description("Note body / text.")] public required string Body  { get; init; }
    }

    public sealed class Output { public Guid Id { get; init; } }
}

Throwing is fine. The runner catches the exception and returns the error to the model, which can retry with different arguments or fall back, controlled by ToolExceptionPolicy.

Server context a tool can trust (and the model cannot forge)

Some tools must act on data that has to come from the server — the signed-in user's id, the tenant, a permission scope. Putting it in the tool's input is unsafe: input is part of the schema, so the model fills it in and could send anything. Inherit ToolBase<TScope, TInput, TOutput> instead — TScope comes first (trusted server data), then the model's TInput. The caller supplies the value per call via context:; it never reaches the model, so it cannot be read or forged through the prompt and flows untouched through the whole pipeline.

public sealed record UserContext(Guid UserId, string UserName, Guid TenantId);

public sealed class GetMyOrdersTool(IOrderRepository orders)
    : ToolBase<UserContext, GetMyOrdersTool.Input, GetMyOrdersTool.Output>
{
    public override string Name        => "get_my_orders";
    public override string Description => "Lists the signed-in user's orders.";

    // context = trusted server data (first); input = model arguments (second).
    public override async Task<Output> ExecuteAsync(UserContext context, Input input, CancellationToken ct)
    {
        var rows = await orders.GetForUserAsync(context.UserId, input.Status, ct);
        return new Output { Count = rows.Count };
    }

    public sealed class Input  { [Description("Optional status filter.")] public string? Status { get; init; } }
    public sealed class Output { public int Count { get; init; } }
}

Supply the context on the builder with .WithContext(...) — each scoped tool resolves the value matching its TScope by type:

var user = new UserContext(currentUser.Id, currentUser.Name, currentUser.TenantId);

await ai.Agent(new GPT5(), prompt)
    .AddTool<GetMyOrdersTool>()
    .WithContext(user)                // call .WithContext again per extra scoped context type
    .RunAsync();

The runner guarantees context is non-null and correctly typed before calling, so you never null-check a missing context. If a scoped tool runs but no matching value was supplied, the runner throws AiToolContextException to you (a wiring mistake caught at first run), never to the model. Validate the context's contents (permissions, etc.) inside the tool as usual.

Sub-agents: delegate to a specialist

A sub-agent is the agent-level counterpart of a tool: a named specialist with its own model, own tools and own prompt that a parent run delegates to. Inherit AgentBase<TOutput> (chat-driven — it reads the forwarded conversation) or AgentBase<TInput, TOutput> (parametrized — the parent model fills TInput, and the Prompt is a Scriban template that input fills). List its tools typeof-free with Toolset.Of<…>().

public sealed class ConversionAgent : AgentBase<string>
{
    public override string Name        => "conversion";
    public override string Description => "Onboards a customer onto the exchange: sign-up, KYC, first deposit.";
    public override IAgentLlm Llm      => new Grok43 { MaxTokens = 8_000 };          // its own model
    public override string Prompt      => "You onboard the customer onto the exchange. ...";
    public override IReadOnlyList<Type> Tools => Toolset.Of<GenerateLinkTool, ContactSaveTool>();
}

builder.Services.AddAiAgent<ConversionAgent>();   // register it (and AddAiTools<T>() for its tools)

Expose sub-agents on any run with .AddAgent<T>(). A common shape is a cheap router/persona model that delegates to specialists and re-voices the reply — so the persona and language stay in one place:

var reply = await ai.Chat(new Haiku45(), routerPrompt, history)
    .AddAgent<ConversionAgent>()       // each specialist: own model + own tools + own prompt
    .AddAgent<SupportAgent>()
    .AddAgent<AnalysisAgent>()
    .WithContext(customer)             // forwarded down to each sub-agent's scoped tools
    .RunAsync();

The router picks a specialist by its Description; the specialist runs its own loop and returns text the parent re-voices. .WithContext(...) and — under a chat parent — the conversation are forwarded to the sub-agent (set ForwardChat => false to isolate it). The sub-agent's cost rolls into Total and NestedAiCalls. Full guide: Instruction/subagents.md.

Registering tools and MCP servers

Add tools on the builder (above), or register global defaults. Globally registered defaults are opt-in — a call gets them only when it asks, with .AddDefaultTools() / .AddDefaultMcp(). This keeps a tool registered for one flow from silently leaking into every other agent call.

builder.Services.AddAiTools<SaveNoteTool>();          // type, resolved from DI
builder.Services.AddAiTools(new ReportBugTool());     // pre-built instance
builder.Services.AddAiMcp(new Mcp("github", "https://mcp.example.com/sse", token));

// opt a specific call into the registered set:
await ai.Agent(new GPT5(), prompt).AddDefaultTools().AddDefaultMcp().RunAsync();

External MCP (client only)

Mcp describes a remote Model Context Protocol server reached over HTTPS/SSE. Its tools are presented to the model as "{server}.{tool}" (for example github.read_file); AllowedTools whitelists which ones are exposed.

var mcp = new Mcp(
    name:  "github",
    url:   "https://mcp.example.com/sse",
    token: bearer,                                  // optional Authorization: Bearer
    allowedTools: ["read_file", "create_issue"]);   // optional whitelist

Per-call options

Every knob is a chainable builder method:

await ai.Agent(new GPT5(), prompt)
    .AddTool<SaveNoteTool>()
    .MaxIterations(12)
    .MaxParallelToolCalls(8)
    .Timeout(TimeSpan.FromMinutes(2))
    .AllowOnly("save_note", "github.read_file")
    .OnToolCall(async (call, ct) => call.Name != "delete_everything")   // veto hook
    .RunAsync();

Method	Purpose
.MaxIterations(n)	Hard ceiling on agent turns
.MaxParallelToolCalls(n)	Concurrency for tool execution within a turn (surplus is queued, never dropped)
.Timeout(t)	Wall-clock limit for the whole run
.AllowOnly(names…)	Per-call allow-list of tool names
.OnToolCall((call, ct) => …)	Async hook before each tool; return false to block it
.AddDefaultTools() / .AddDefaultMcp()	Opt IN to DI-registered defaults (off otherwise)
.MaxNestedDepth(n)	Bound on agent-to-tool-to-agent nesting

The audit trail: ResultAgent<T>

ResultAgent<T> extends Result<T> with everything that happened during the run, for logging, debugging, replay or cross-model verification:

public class ResultAgent<T> : Result<T>
{
    public int                          Iterations    { get; }  // model round-trips
    public IReadOnlyList<ToolInvocation> ToolCalls     { get; }  // ordered, with timings
    public AiUsageSummary               Request       { get; }  // this agent's own turns
    public AiUsageSummary               Total         { get; }  // whole run, including nested AI
    public AiUsageScope?                Usage         { get; }  // full call tree
    public IReadOnlyList<AiUsageScope>  NestedAiCalls { get; }  // flat list of nested calls
}

foreach (var call in result.ToolCalls)
    Console.WriteLine($"[{call.Iteration}] {call.Name}: {call.Duration.TotalMilliseconds:F0} ms"
                    + (call.IsError ? $" failed: {call.Error}" : ""));

Each ToolInvocation records the iteration, name, input, output, error and error type, duration, originating MCP server, whether it was blocked, and any nested AI usage the tool itself consumed.

Nested cost tracking

If a tool injects IAiProvider and calls another model, that cost is tracked too. This is the difference between Request and Total:

• Request: only the main agent's own model turns.
• Total: the entire run, including every nested model call made inside tools and sub-agents. This is your end-to-end number for billing or quota.

NestedAiCalls gives the flat list (model, tokens, cost, duration, originating tool) and Usage the full tree. Tracking flows with the async context, so concurrent requests stay isolated with no plumbing on your side.

Streaming an agent run

.RunStreamAsync() emits a sealed AgentEvent hierarchy so you can drive a live UI:

await foreach (var evt in ai.Agent(new Sonnet46(), prompt).AddTool<SaveNoteTool>().RunStreamAsync())
{
    switch (evt)
    {
        case AgentIterationStartedEvent e:   ui.NewTurn(e.Iteration);                break;
        case AgentToolCallStartedEvent s:    ui.ShowTool(s.ToolName, s.CallId);      break;
        case AgentToolCallCompletedEvent d:  ui.MarkDone(d.Invocation);              break;
        case AgentFinalTextEvent f:          ui.AppendFinal(f.Text);                 break;
        case AgentCompletedEvent<MyAnswer> c: ui.Done(c.Result);                     break;
        case AgentFailedEvent x:             ui.Error(x.Error);                       break;
    }
}

ai.Chat(llm, prompt, history).RunStreamAsync() does the same resuming from a prior transcript with full tool-calling (needs an IAgentLlm model).

Chat and agent API at a glance

Call	Returns	Tools	Streaming
ChatAsync(llm, prompt, history)	Result<T>	no	no
ChatStreamAsync(llm, prompt, history)	IAsyncEnumerable<string>	no	tokens
Agent(agentLlm, prompt).RunAsync()	ResultAgent<T>	yes	no
Agent(agentLlm, prompt).RunStreamAsync()	IAsyncEnumerable<AgentEvent>	yes	events
Chat(llm, prompt, history).RunAsync()	Result<T>	yes	no
Chat(llm, prompt, history).RunStreamAsync()	IAsyncEnumerable<AgentEvent>	yes	events

Every entry point has a plain-string overload for when you do not need a typed response. On either builder, .AddAgent<T>() delegates to a sub-agent — a specialist with its own model, tools and prompt.

---

Ready-made prompts

Zonit.Extensions.Ai.Prompts is a separate package of reusable, production-grade prompt templates. When the task matches one, install the package and use it instead of writing a prompt from scratch.

dotnet add package Zonit.Extensions.Ai.Prompts

TranslatePrompt translates text as a native writer would, applying per-language localization rules (quotation marks, number and date formats, dash conventions and register) for 19 languages (en, pl, de, es, fr, it, pt, nl, sv, da, no, fi, ru, uk, cs, sk, hu, tr, ar), with a general fallback for any other target.

using Zonit.Extensions.Ai.Prompts;

var result = await ai.GenerateAsync(
    new GPT5(),
    new TranslatePrompt { Content = "Hello world!", Target = "pl" });

string translated = result.Value;   // "Witaj świecie!" — translated text as a plain string

More templates will be added over time. Details in Instruction/prompt-library.md.

---

Cost tracking and estimation

Read actual cost from any result:

var r = await ai.GenerateAsync(new GPT5(), prompt);
Console.WriteLine($"{r.MetaData.InputTokens} in / {r.MetaData.OutputTokens} out");
Console.WriteLine($"Total: {r.MetaData.TotalCost}");

Or estimate before you call:

Price perCall   = ai.CalculateCost(new GPT5(), inputTokens: 1_000, outputTokens: 500);
Price embedCost = ai.CalculateCost(new TextEmbedding3Large(), inputTokens: 1_000);
Price imageCost = ai.CalculateCost(new GPTImage15 { Quality = GPTImage15.QualityType.High,
                                                    Size = GPTImage15.SizeType.Square });
Price estimate  = ai.EstimateCost(new GPT5(), "your prompt text...", estimatedOutputTokens: 500);

---

Configuration

Hierarchy

Everything lives under the Ai section: shared settings such as Resilience at the top, one subsection per provider.

{
  "Ai": {
    "Resilience": { "MaxRetryAttempts": 3, "AttemptTimeout": "00:10:00" },
    "OpenAi":     { "ApiKey": "sk-...", "OrganizationId": "org-...", "Timeout": "00:15:00" },
    "Anthropic":  { "ApiKey": "sk-ant-..." },
    "Agent":      { "MaxIterations": 100, "MaxParallelToolCalls": 16 }
  }
}

Code-based overrides

appsettings.json is bound first; code overrides are applied on top through PostConfigure, so you set only what you need:

// API key inline
builder.Services.AddAiOpenAi("sk-...");

// Or full options (for example an Azure OpenAI endpoint)
builder.Services.AddAiOpenAi(o =>
{
    o.OrganizationId = "org-...";
    o.BaseUrl        = "https://my-azure-openai.openai.azure.com";
    o.Timeout        = TimeSpan.FromMinutes(15);   // per-provider override
});

Keep secrets in User Secrets, environment variables or a vault; do not hard-code keys in source. Each provider extends a common AiProviderOptions (ApiKey, BaseUrl, Timeout) and adds its own fields where relevant.

---

Resilience

Built on Microsoft.Extensions.Http.Resilience, tuned for long-running AI requests. Retry, timeout and circuit-breaker apply to every provider and are configurable under Ai:Resilience (or in code via AddAi(o => o.Resilience...)).

Setting	Default	Meaning
TotalRequestTimeout	40 min	Budget for the whole pipeline, including retries
AttemptTimeout	10 min	Timeout for a single attempt
MaxRetryAttempts	3	Retries on transient failures
RetryBaseDelay / RetryMaxDelay	2 s / 30 s	Exponential backoff bounds
UseJitter	true	Randomises delays to avoid thundering herds
CircuitBreakerFailureRatio	0.5	Failure ratio that opens the circuit

Requests are retried on network errors, timeouts, HTTP 429, and 5xx responses.

---

AOT and trimming

The library targets net10.0 and is built for Native AOT and trimming. For the documented typed-prompt pattern the structured-output path is fully source-generated and runs zero reflection: incremental generators emit, at build time, the request JSON Schema and the JsonTypeInfo<T> for T and its whole graph, plus the Scriban property binding. Each provider also serializes its wire DTOs through its own source-generated JsonSerializerContext.

Path	AOT / trim
Text · streaming · embeddings · images · transcription	✅ source-generated, zero reflection
Structured output via PromptBase<T>	✅ schema and JsonTypeInfo<T> generated at build time
Scriban prompt rendering	✅ source-generated property binding
Tools · agents · MCP	⚠️ require dynamic code; tool arguments are (de)serialized reflectively, and these APIs are annotated [RequiresUnreferencedCode] / [RequiresDynamicCode]
Ad-hoc response types the generator never sees (for example SimplePrompt<T> over a type shaped at runtime)	⚠️ reflection fallback; not AOT-safe

The only reflection on the structured-output path is a gated fallback for response types the source generator did not see; for the typed-prompt pattern it never runs, so the trim and AOT analyzers stay clean without suppressing real warnings.

To publish Native AOT, set <PublishAot>true</PublishAot> in your app and:

dotnet publish -c Release -r win-x64

---

Provider packages

Install only what you use. Every package version and download count below is live from NuGet.

Core

Package	Version	Downloads
Zonit.Extensions.Ai
Zonit.Extensions.Ai.Abstractions
Zonit.Extensions.Ai.Prompts

Providers

Provider	Package	Version	Downloads
OpenAI	Zonit.Extensions.Ai.OpenAi
Anthropic	Zonit.Extensions.Ai.Anthropic
Google	Zonit.Extensions.Ai.Google
xAI	Zonit.Extensions.Ai.X
DeepSeek	Zonit.Extensions.Ai.DeepSeek
Mistral	Zonit.Extensions.Ai.Mistral
Groq	Zonit.Extensions.Ai.Groq
Together AI	Zonit.Extensions.Ai.Together
Fireworks	Zonit.Extensions.Ai.Fireworks
Cohere	Zonit.Extensions.Ai.Cohere
Perplexity	Zonit.Extensions.Ai.Perplexity
Alibaba	Zonit.Extensions.Ai.Alibaba
Baidu	Zonit.Extensions.Ai.Baidu
Zhipu	Zonit.Extensions.Ai.Zhipu
Moonshot	Zonit.Extensions.Ai.Moonshot
01.AI	Zonit.Extensions.Ai.Yi

Each provider is registered with its matching extension method: AddAiOpenAi(), AddAiAnthropic(), AddAiGoogle(), AddAiX(), AddAiGroq(), and so on.

---

Architecture

The library follows Clean Architecture and SOLID principles:

• Abstractions-only domain. Zonit.Extensions.Ai.Abstractions holds every interface and contract (IAiProvider, ILlm, IPrompt, ITool, results, agent events). Your domain layer can define prompts and tools referencing only this package, with no HTTP, DI or Scriban.
• Self-contained providers. Each provider is an independent package with its own options, DI extension, model classes and JsonSerializerContext. Adding a provider never touches the core (Open/Closed).
• Idempotent registration. AddAi* methods use TryAdd* and TryAddEnumerable, so they are safe to call from multiple modules or plugins.
• The core owns the agent loop. Providers implement a thin IAgentProviderAdapter; iteration, parallel tool execution, MCP merging and audit assembly live in the core.

Zonit.Extensions.Ai.Abstractions    interfaces and contracts (domain-safe)
Zonit.Extensions.Ai                 orchestrator, agent runtime, MCP client, templating, DI
Zonit.Extensions.Ai.SourceGenerators  registration, schema and binding generators
Zonit.Extensions.Ai.{Provider}      one package per provider
Zonit.Extensions.Ai.Prompts         optional ready-made prompts

---

Requirements and license

• .NET 10.0 (single target framework).
• Released under the MIT License. See LICENSE.txt.

Issues and pull requests are welcome at github.com/Zonit/Zonit.Extensions.Ai.