feat: vastly simplify providers

README.md

@@ -26,12 +26,13 @@ Models and Adapters for these providers are shipped out of the box:
 As well as adapters for commonly supported [Open Responses](https://www.openresponses.org/) and
 [OpenAI Completions](https://developers.openai.com/api/docs/guides/completions) API's so you can easily BYO models.
 
-Provider model constructors accept options like `apiKey` and `baseUrl`. If you do not pass an API key explicitly,
+Provider model factories accept options like `apiKey` and `baseUrl`. If you do not pass an API key explicitly,
 provider-specific environment variables are used where supported.
 
 ```ts
 import z from "npm:zod";
-import { Agent, OpenAIModel, Tool } from "jsr:@alphaxiv/agents";
+import { Agent, Tool } from "jsr:@alphaxiv/agents";
+import { openAIModel } from "jsr:@alphaxiv/agents/openai";
 
 const calculator = new Tool({
   name: "calculator",
@@ -50,7 +51,7 @@ const calculator = new Tool({
 });
 
 const agent = new Agent({
-  model: new OpenAIModel({ model: "gpt-4.1-mini" }), // Or via shorthand "openai:gpt-4.1-mini"
+  model: openAIModel({ model: "gpt-4.1-mini" }), // Or via shorthand "openai:gpt-4.1-mini"
   instructions: "You are a helpful assistant. Use the calculator when math is needed.",
   tools: [calculator],
 });
@@ -77,10 +78,11 @@ If you provide an `output` Zod schema, the agent guarantees that the output will
 
 ```ts
 import z from "npm:zod";
-import { Agent, GeminiModel } from "jsr:@alphaxiv/agents";
+import { Agent } from "jsr:@alphaxiv/agents";
+import { geminiModel } from "jsr:@alphaxiv/agents/gemini";
 
 const paperExtractor = new Agent({
-  model: new GeminiModel({ model: "gemini-2.0-flash" }),
+  model: geminiModel({ model: "gemini-2.0-flash" }),
   instructions: "Extract the paper title and abstract from the raw text.",
   output: z.object({
     title: z.string(),
@@ -111,7 +113,8 @@ Tool names are normalized internally so provider-specific restrictions do not le
 
 ```ts
 import z from "npm:zod";
-import { Agent, AnthropicModel, Tool } from "jsr:@alphaxiv/agents";
+import { Agent, Tool } from "jsr:@alphaxiv/agents";
+import { anthropicModel } from "jsr:@alphaxiv/agents/anthropic";
 
 const getImage = new Tool({
   name: "get_image",
@@ -131,7 +134,7 @@ const getImage = new Tool({
 });
 
 const agent = new Agent({
-  model: new AnthropicModel({ model: "claude-haiku-4-5" }),
+  model: anthropicModel({ model: "claude-haiku-4-5" }),
   instructions: "You are a helpful assistant.",
   tools: [getImage],
 });
@@ -143,7 +146,8 @@ If a tool already knows the final answer, return `new ModelOutput(...)` to end t
 
 ```ts
 import z from "npm:zod";
-import { Agent, ModelOutput, OpenAIModel, Tool } from "jsr:@alphaxiv/agents";
+import { Agent, ModelOutput, Tool } from "jsr:@alphaxiv/agents";
+import { openAIModel } from "jsr:@alphaxiv/agents/openai";
 
 const reportHelpfulIds = new Tool({
   name: "report_helpful_ids",
@@ -155,7 +159,7 @@ const reportHelpfulIds = new Tool({
 });
 
 const agent = new Agent({
-  model: new OpenAIModel({ model: "gpt-4.1-mini" }),
+  model: openAIModel({ model: "gpt-4.1-mini" }),
   instructions: "Find relevant ids, then call report_helpful_ids.",
   tools: [reportHelpfulIds],
 });
@@ -173,12 +177,14 @@ type that occured and will automatically determine whether to retry, fallback to
 differently.
 
 ```ts
-import { Agent, AnthropicModel, OpenAIModel } from "jsr:@alphaxiv/agents";
+import { Agent } from "jsr:@alphaxiv/agents";
+import { anthropicModel } from "jsr:@alphaxiv/agents/anthropic";
+import { openAIModel } from "jsr:@alphaxiv/agents/openai";
 
 const agent = new Agent({
   model: [
-    new OpenAIModel({ model: "gpt-4.1-mini" }),
-    new AnthropicModel({ model: "claude-haiku-4-5" }),
+    openAIModel({ model: "gpt-4.1-mini" }),
+    anthropicModel({ model: "claude-haiku-4-5" }),
   ],
   instructions: "You are a reliable assistant.",
 });
@@ -191,8 +197,8 @@ You can customize the retry strategy to match your needs:
 ```ts
 const agent = new Agent({
   model: [
-    new OpenAIModel({ model: "gpt-5.4-mini" }),
-    new AnthropicModel({ model: "claude-haiku-4-5" }),
+    openAIModel({ model: "gpt-5.4-mini" }),
+    anthropicModel({ model: "claude-haiku-4-5" }),
   ],
   instructions: "You are a helpful assistant.",
   retryStrategy: {
@@ -237,10 +243,11 @@ Use `agent.stream(...)` when you want token-by-token output, reasoning deltas, a
 
 ```ts
 import process from "node:process";
-import { Agent, OpenAIModel } from "jsr:@alphaxiv/agents";
+import { Agent } from "jsr:@alphaxiv/agents";
+import { openAIModel } from "jsr:@alphaxiv/agents/openai";
 
 const agent = new Agent({
-  model: new OpenAIModel({ model: "gpt-5.4-mini" }),
+  model: openAIModel({ model: "gpt-5.4-mini" }),
   instructions: "You are a helpful assistant.",
 });
 
@@ -268,7 +275,8 @@ Built-in way to convert agents into a CLI app.
 
 ```ts
 import z from "npm:zod";
-import { Agent, cli, OpenAIModel, Tool } from "jsr:@alphaxiv/agents";
+import { Agent, cli, Tool } from "jsr:@alphaxiv/agents";
+import { openAIModel } from "jsr:@alphaxiv/agents/openai";
 
 const echo = new Tool({
   name: "echo",
@@ -278,7 +286,7 @@ const echo = new Tool({
 });
 
 const agent = new Agent({
-  model: new OpenAIModel({ model: "gpt-4.1-mini" }),
+  model: openAIModel({ model: "gpt-4.1-mini" }),
   instructions: "You are a helpful assistant.",
   tools: [echo],
 });
@@ -299,10 +307,11 @@ The built-in CLI supports:
 Pass files directly as chat items.
 
 ```ts
-import { Agent, OpenRouterModel } from "jsr:@alphaxiv/agents";
+import { Agent } from "jsr:@alphaxiv/agents";
+import { openrouterModel } from "jsr:@alphaxiv/agents/openrouter";
 
 const agent = new Agent({
-  model: new OpenRouterModel({ model: "openrouter:meta-llama/llama-4-maverick" }),
+  model: openrouterModel({ model: "meta-llama/llama-4-maverick" }),
   instructions: "You are a helpful assistant.",
 });
 

deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@alphaxiv/agents",
-  "version": "0.5.11",
+  "version": "0.6.0",
   "license": "MIT",
   "fmt": {
     "lineWidth": 120
@@ -10,31 +10,17 @@
     "./cli": "./src/cli.ts",
     "./client": "./src/client.ts",
 
-    "./adapters/anthropic": "./src/adapters/anthropic/adapter.ts",
-    "./anthropic": "./src/adapters/anthropic/model.ts",
+    "./anthropic": "./src/adapters/anthropic/adapter.ts",
+    "./google-genai": "./src/adapters/google_genai/adapter.ts",
+    "./gemini": "./src/adapters/gemini/adapter.ts",
+    "./vertex-ai": "./src/adapters/vertex_ai/adapter.ts",
+    "./openai_completions": "./src/adapters/openai_completions/adapter.ts",
+    "./open_responses": "./src/adapters/open_responses/adapter.ts",
+    "./openrouter": "./src/adapters/openrouter/adapter.ts",
+    "./tributary": "./src/adapters/tributary/adapter.ts",
+    "./openai": "./src/adapters/openai/adapter.ts",
+    "./sid": "./src/adapters/sid/adapter.ts",
 
-    "./adapters/google-genai": "./src/adapters/google_genai/adapter.ts",
-
-    "./adapters/gemini": "./src/adapters/gemini/adapter.ts",
-    "./gemini": "./src/adapters/gemini/model.ts",
-
-    "./adapters/vertex-ai": "./src/adapters/vertex_ai/adapter.ts",
-    "./vertex-ai": "./src/adapters/vertex_ai/model.ts",
-
-    "./adapters/openai_completions": "./src/adapters/openai_completions/adapter.ts",
-    "./adapters/open_responses": "./src/adapters/open_responses/adapter.ts",
-
-    "./adapters/openrouter": "./src/adapters/openrouter/adapter.ts",
-    "./openrouter": "./src/adapters/openrouter/model.ts",
-
-    "./adapters/tributary": "./src/adapters/tributary/adapter.ts",
-    "./tributary": "./src/adapters/tributary/model.ts",
-
-    "./adapters/openai": "./src/adapters/openai/adapter.ts",
-    "./openai": "./src/adapters/openai/model.ts",
-
-    "./adapters/sid": "./src/adapters/sid/adapter.ts",
-    "./sid": "./src/adapters/sid/model.ts",
     "./sid/tools": "./src/adapters/sid/tools.ts"
   },
   "tasks": {
@@ -55,5 +41,14 @@
     "@std/uuid": "jsr:@std/uuid@^1.1.0",
     "openai": "npm:openai@^6.33.0",
     "zod": "npm:zod@^4.3.6"
+  },
+  "publish": {
+    "include": [
+      "LICENSE",
+      "README.md",
+      "mod.ts",
+      "src/**/*.ts",
+      "tests/**/*.ts"
+    ]
   }
 }

examples/custom-providers.ts

@@ -0,0 +1,47 @@
+import { Agent } from "@alphaxiv/agents";
+import { openResponsesModel } from "@alphaxiv/agents/open_responses";
+import { openAICompletionsModel } from "@alphaxiv/agents/openai_completions";
+
+// OpenAI chat completions
+{
+  const model = openAICompletionsModel({
+    model: "qwen/qwen3-coder-next",
+    openAIOptions: {
+      baseURL: "http://localhost:1234/v1",
+      apiKey: "n/a",
+    },
+  });
+
+  const agent = new Agent({
+    model,
+    instructions: "you say the number 8",
+  });
+
+  const result = await agent.run(
+    `On a scale of 1 to 10 how would you rate the JSR package @alphaxiv/agents`,
+  );
+
+  console.log(result);
+}
+
+// For LMStudio, prefer open responses so you get reasoning
+{
+  const model = openResponsesModel({
+    model: "qwen/qwen3-coder-next",
+    openAIOptions: {
+      baseURL: "http://localhost:1234/v1",
+      apiKey: "n/a",
+    },
+  });
+
+  const agent = new Agent({
+    model,
+    instructions: "you say the number 8",
+  });
+
+  const result = await agent.run(
+    `On a scale of 1 to 10 how would you rate the JSR package @alphaxiv/agents`,
+  );
+
+  console.log(result);
+}

examples/tracing.ts

@@ -10,15 +10,15 @@ import process from "node:process";
 import z from "zod";
 import {
   Agent,
-  AnthropicModel,
   type ChatItem,
-  OpenAIModel,
   type PartialTraceEvent,
   registerGlobalTracer,
   Tool,
   type TraceEvent,
   withTrace,
 } from "../mod.ts";
+import { anthropicModel } from "@alphaxiv/agents/anthropic";
+import { openAIModel } from "@alphaxiv/agents/openai";
 
 const snackMenu = [
   { name: "miso ramen cup", calories: 420, prepMinutes: 6, price: 8 },
@@ -70,8 +70,8 @@ async function main() {
   const subagent = new Agent({
     name: "Snack Specialist",
     model: [
-      new OpenAIModel({ model: "gpt-5.4-nano" }),
-      new AnthropicModel({ model: "claude-sonnet-4-6", thinkingLevel: "adaptive" }),
+      openAIModel({ model: "gpt-5.4-nano" }),
+      anthropicModel({ model: "claude-sonnet-4-6", thinkingLevel: "adaptive" }),
     ],
     instructions: [
       "You are the snack specialist.",
@@ -101,7 +101,7 @@ async function main() {
 
   const agent = new Agent({
     name: "Main Agent",
-    model: new AnthropicModel({ model: "claude-sonnet-4-6", thinkingLevel: "adaptive" }),
+    model: anthropicModel({ model: "claude-sonnet-4-6", thinkingLevel: "adaptive" }),
     instructions: [
       "You coordinate requests for a snack specialist.",
       "Always call delegate_specialist exactly once before answering.",

mod.ts

@@ -21,17 +21,12 @@ export type { CliIo, CliOptions } from "./src/cli.ts";
 export { ModelOutput, Tool } from "./src/tool.ts";
 export type { AnyTool, ExecuteContext, ExecuteFunc, ExecuteFuncInput, ExecuteResult } from "./src/tool.ts";
 
-export { Adapter } from "./src/adapters/adapter.ts";
+export type { Adapter } from "./src/adapters/adapter.ts";
 export type { AdapterOptions, AdapterStreamOptions } from "./src/adapters/adapter.ts";
 
-export { Model } from "./src/adapters/model.ts";
-export type { ModelOptions } from "./src/adapters/model.ts";
-
 export { resolveModel } from "./src/adapters/model_resolver.ts";
-export type { ModelLike, ModelString } from "./src/adapters/model_resolver.ts";
+export type { AdapterLike, ModelString } from "./src/adapters/model_resolver.ts";
 
-export { AnthropicModel } from "./src/adapters/anthropic/model.ts";
-export type { AnthropicModelOptions } from "./src/adapters/anthropic/model.ts";
 export type {
   AnthropicModels,
   EffortLevel,
@@ -42,32 +37,16 @@ export type {
   ThinkingLevel,
 } from "./src/adapters/anthropic/models.ts";
 
-export { GeminiModel } from "./src/adapters/gemini/model.ts";
-export type { GeminiModelOptions } from "./src/adapters/gemini/model.ts";
-
-export { OpenAIModel } from "./src/adapters/openai/model.ts";
-export type { OpenAIModelOptions } from "./src/adapters/openai/model.ts";
 export type {
   OpenAIModelModality,
   OpenAIModels,
   OpenAIReasoningEffort,
   SupportedReasoningEffort,
 } from "./src/adapters/openai/models.ts";
 
-export { OpenRouterModel } from "./src/adapters/openrouter/model.ts";
-export type { OpenRouterModelOptions } from "./src/adapters/openrouter/model.ts";
 export type { OpenRouterModels, OpenRouterReasoningEffort } from "./src/adapters/openrouter/models.ts";
 
 export type { SidModels } from "./src/adapters/sid/adapter.ts";
-export { SidModel } from "./src/adapters/sid/model.ts";
-export type { SidModelOptions } from "./src/adapters/sid/model.ts";
-
-export { TributaryModel } from "./src/adapters/tributary/model.ts";
-export type { TributaryModelOptions } from "./src/adapters/tributary/model.ts";
-export type { TributaryModels } from "./src/adapters/tributary/models.ts";
-
-export { VertexAiModel } from "./src/adapters/vertex_ai/model.ts";
-export type { VertexAiModelOptions, VertexAiModelPriority } from "./src/adapters/vertex_ai/model.ts";
 
 export { SidEmbeddingSearchTool, SidReportHelpfulIdsTool, SidTextSearchTool } from "./src/adapters/sid/tools.ts";
 

src/adapters/adapter.ts

@@ -23,16 +23,13 @@ export interface AdapterStreamOptions<zO, zI> {
   signal: AbortSignal;
 }
 
-export abstract class Adapter<SupportedModels extends string> {
-  abstract name: string;
-  model: SupportedModels;
-
-  constructor(options: AdapterOptions<SupportedModels>) {
-    this.model = options.model;
-  }
-
-  abstract stream<zO, zI>(options: AdapterStreamOptions<zO, zI>): AdapterStreamIterator;
-
+export type Adapter<zO, zI> = {
+  /** Name of the provider */
+  provider: string;
+  /** Name of the model on the provider */
+  model: string;
+  /** Function to actually stream the results */
+  stream: (options: AdapterStreamOptions<zO, zI>) => AdapterStreamIterator;
   /**
    * Classify an error using provider-specific error types.
    * Override this in adapter implementations to provide precise error classification.
@@ -41,5 +38,5 @@ export abstract class Adapter<SupportedModels extends string> {
    * @param error The error thrown during a model call
    * @returns ClassifiedError if the adapter can classify it, null to fall back to heuristics
    */
-  classifyError?(error: unknown): ClassifiedError | null;
-}
+  classifyError?: (error: unknown) => ClassifiedError | null;
+};