feat(sync): add CrofAI catalog sync

package.json

@@ -23,6 +23,7 @@
     "deepinfra:sync": "bun ./packages/core/script/sync-models.ts deepinfra",
     "cloudflare:sync": "bun ./packages/core/script/sync-models.ts cloudflare-workers-ai",
     "chutes:sync": "bun ./packages/core/script/sync-models.ts chutes",
+    "crof:sync": "bun ./packages/core/script/sync-models.ts crof",
     "databricks:generate": "bun ./packages/core/script/generate-databricks.ts",
     "helicone:generate": "bun ./packages/core/script/generate-helicone.ts",
     "huggingface:sync": "bun ./packages/core/script/sync-models.ts huggingface",

packages/core/src/sync/index.ts

@@ -11,6 +11,7 @@ import { baseten } from "./providers/baseten.js";
 import { chutes } from "./providers/chutes.js";
 import { cloudflareWorkersAi } from "./providers/cloudflare-workers-ai.js";
 import { cortecs } from "./providers/cortecs.js";
+import { crof } from "./providers/crof.js";
 import { crossmodel } from "./providers/crossmodel.js";
 import { deepinfra } from "./providers/deepinfra.js";
 import { digitalocean } from "./providers/digitalocean.js";
@@ -120,6 +121,7 @@ export const providers: {
   chutes: SyncProvider<any>;
   "cloudflare-workers-ai": SyncProvider<any>;
   cortecs: SyncProvider<any>;
+  crof: SyncProvider<any>;
   crossmodel: SyncProvider<any>;
   deepinfra: SyncProvider<any>;
   digitalocean: SyncProvider<any>;
@@ -151,6 +153,7 @@ export const providers: {
   chutes,
   "cloudflare-workers-ai": cloudflareWorkersAi,
   cortecs,
+  crof,
   crossmodel,
   deepinfra,
   digitalocean,
@@ -194,7 +197,7 @@ export const groups = {
     "vercel",
   ],
   cloudflare: ["cloudflare-workers-ai"],
-  direct: ["ambient", "anthropic", "baseten", "chutes", "cortecs", "deepinfra", "digitalocean", "google", "hyper", "openai", "ovhcloud", "pioneer", "tinfoil", "venice", "wandb", "xai"],
+  direct: ["ambient", "anthropic", "baseten", "chutes", "cortecs", "crof", "deepinfra", "digitalocean", "google", "hyper", "openai", "ovhcloud", "pioneer", "tinfoil", "venice", "wandb", "xai"],
 } as const;
 
 type ProviderID = keyof typeof providers;

packages/core/src/sync/providers/crof.ts

@@ -0,0 +1,126 @@
+import { z } from "zod";
+
+import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
+import { factorBaseModel } from "./openrouter.js";
+
+const API_ENDPOINT = "https://crof.ai/v1/models";
+
+const Price = z.string().regex(/^\d+(?:\.\d+)?$/).transform(Number);
+
+export const CrofModel = z.object({
+  id: z.string().min(1),
+  name: z.string().min(1),
+  created: z.number().int().nonnegative(),
+  context_length: z.number().int().positive(),
+  max_completion_tokens: z.number().int().positive(),
+  custom_reasoning: z.boolean(),
+  reasoning_effort: z.boolean().optional(),
+  pricing: z.object({
+    prompt: Price,
+    completion: Price,
+    cache_prompt: Price.optional(),
+    discount: z.number().nonnegative().max(100).optional(),
+  }).passthrough(),
+}).passthrough();
+
+export const CrofResponse = z.object({
+  data: z.array(CrofModel),
+}).passthrough();
+
+export type CrofModel = z.infer<typeof CrofModel>;
+
+export const crof = {
+  id: "crof",
+  name: "CrofAI",
+  modelsDir: "providers/crof/models",
+  skipCreates: true,
+  deleteMissing: true,
+  sourceID(model) {
+    return model.id;
+  },
+  skippedNotice(ids) {
+    if (ids.length === 0) return [];
+    return [
+      `${ids.length} CrofAI models returned by the API were not created because the endpoint does not provide authoritative modalities, model capabilities, or descriptions. Existing models are still fully synchronized.`,
+      `Skipped remote IDs: ${ids.map((id) => `\`${id}\``).join(", ")}`,
+    ];
+  },
+  async fetchModels() {
+    const response = await fetch(API_ENDPOINT);
+    if (!response.ok) {
+      throw new Error(`CrofAI models request failed: ${response.status} ${response.statusText}`);
+    }
+    return response.json();
+  },
+  parseModels(raw) {
+    return CrofResponse.parse(raw).data;
+  },
+  translateModel(model, context) {
+    const existing = context.existing(model.id);
+    if (existing === undefined) return undefined;
+    return {
+      id: model.id,
+      model: buildCrofModel(model, existing, context.authored(model.id)),
+    };
+  },
+} satisfies SyncProvider<CrofModel>;
+
+export function buildCrofModel(
+  model: CrofModel,
+  existing: ExistingModel,
+  authored: ExistingModel | undefined,
+): SyncedModel {
+  const required = {
+    name: existing.name,
+    description: existing.description,
+    release_date: existing.release_date,
+    last_updated: existing.last_updated,
+    attachment: existing.attachment,
+    temperature: existing.temperature,
+    tool_call: existing.tool_call,
+    open_weights: existing.open_weights,
+    modalities: existing.modalities,
+  };
+  for (const [field, value] of Object.entries(required)) {
+    if (value === undefined) throw new Error(`CrofAI model ${model.id} is missing local ${field} metadata`);
+  }
+
+  const reasoning = model.custom_reasoning;
+  const limit = {
+    input: existing.limit?.input,
+    context: model.context_length,
+    output: model.max_completion_tokens,
+  };
+  const synced: SyncedFullModel = {
+    name: required.name!,
+    description: required.description!,
+    family: existing.family,
+    release_date: required.release_date!,
+    last_updated: required.last_updated!,
+    attachment: required.attachment!,
+    reasoning,
+    reasoning_options: reasoning ? existing.reasoning_options ?? [] : undefined,
+    temperature: required.temperature!,
+    tool_call: required.tool_call!,
+    structured_output: existing.structured_output,
+    knowledge: existing.knowledge,
+    open_weights: required.open_weights!,
+    status: existing.status,
+    interleaved: reasoning ? existing.interleaved : undefined,
+    cost: {
+      input: model.pricing.prompt,
+      output: model.pricing.completion,
+      cache_read: model.pricing.cache_prompt,
+      cache_write: existing.cost?.cache_write,
+      reasoning: existing.cost?.reasoning,
+      tiers: existing.cost?.tiers,
+    },
+    limit,
+    modalities: required.modalities!,
+    provider: existing.provider,
+  };
+
+  return authored?.base_model === undefined
+    ? synced
+    : factorBaseModel(authored.base_model, synced, limit, authored.base_model_omit);
+}

packages/core/test/sync.test.ts

@@ -11,6 +11,7 @@ import {
   type AnthropicModel,
 } from "../src/sync/providers/anthropic.js";
 import { buildCortecsModel, type CortecsModel } from "../src/sync/providers/cortecs.js";
+import { buildCrofModel, CrofResponse, type CrofModel } from "../src/sync/providers/crof.js";
 import {
   buildCrossModel,
   CrossModelResponse,
@@ -2679,6 +2680,69 @@ test("overrides canonical metadata with Cortecs reasoning support", () => {
   });
 });
 
+test("parses CrofAI string prices as per-million numbers", () => {
+  const parsed = CrofResponse.parse({
+    data: [{
+      id: "glm-5.2",
+      name: "Z.ai: GLM 5.2",
+      created: 1_781_650_191,
+      context_length: 1_000_000,
+      max_completion_tokens: 131_072,
+      custom_reasoning: true,
+      reasoning_effort: true,
+      pricing: {
+        prompt: "0.15",
+        completion: "0.52",
+        cache_prompt: "0.02",
+        discount: 50,
+      },
+    }],
+  });
+
+  expect(parsed.data[0]?.pricing).toMatchObject({
+    prompt: 0.15,
+    completion: 0.52,
+    cache_prompt: 0.02,
+    discount: 50,
+  });
+});
+
+test("updates CrofAI authoritative pricing and limits while preserving local metadata", () => {
+  const model: CrofModel = {
+    id: "glm-5.2",
+    name: "Z.ai: GLM 5.2",
+    created: 1_781_650_191,
+    context_length: 1_000_000,
+    max_completion_tokens: 131_072,
+    custom_reasoning: true,
+    reasoning_effort: true,
+    pricing: { prompt: 0.15, completion: 0.52, cache_prompt: 0.02, discount: 50 },
+  };
+  const existing: ExistingModel = {
+    name: "GLM 5.2",
+    description: "Preserved description",
+    release_date: "2026-06-15",
+    last_updated: "2026-06-15",
+    attachment: false,
+    reasoning: true,
+    reasoning_options: [{ type: "effort", values: ["none", "low", "medium", "high"] }],
+    temperature: true,
+    tool_call: true,
+    open_weights: true,
+    cost: { input: 0.3, output: 1.05, cache_read: 0.05, cache_write: 0 },
+    limit: { context: 202_752, output: 202_752 },
+    modalities: { input: ["text"], output: ["text"] },
+  };
+
+  expect(buildCrofModel(model, existing, undefined)).toMatchObject({
+    name: "GLM 5.2",
+    description: "Preserved description",
+    reasoning_options: [{ type: "effort", values: ["none", "low", "medium", "high"] }],
+    cost: { input: 0.15, output: 0.52, cache_read: 0.02, cache_write: 0 },
+    limit: { context: 1_000_000, output: 131_072 },
+  });
+});
+
 test("syncs OpenRouter reasoning efforts from model metadata", () => {
   const model = buildOpenRouterModel(openRouterModel({
     reasoning: {

providers/crof/models/deepseek-v4-pro-lightning.toml

@@ -1,17 +0,0 @@
-base_model = "deepseek/deepseek-v4-pro"
-name = "DeepSeek V4 Pro Lightning"
-reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high"] }]
-
-[interleaved]
-field = "reasoning_content"
-
-[cost]
-input = 0.8
-output = 1.6
-cache_read = 0.02
-
-[limit]
-output = 131_072
-
-[provider]
-npm = "@ai-sdk/openai-compatible"

providers/crof/models/glm-4.7-flash.toml

@@ -1,16 +0,0 @@
-base_model = "zhipuai/glm-4.7-flash"
-# Crof's request docs establish no model-specific toggle, effort, or budget.
-# https://crof.ai/docs.md (accessed 2026-06-25)
-reasoning_options = []
-
-[cost]
-input = 0.04
-output = 0.3
-cache_read = 0.008
-cache_write = 0
-
-[limit]
-context = 202_752
-
-[provider]
-npm = "@ai-sdk/openai-compatible"

providers/crof/models/glm-4.7.toml

@@ -1,20 +0,0 @@
-base_model = "zhipuai/glm-4.7"
-# Crof's request docs establish no model-specific toggle, effort, or budget.
-# https://crof.ai/docs.md (accessed 2026-06-25)
-reasoning_options = []
-
-[interleaved]
-field = "reasoning_content"
-
-[cost]
-input = 0.25
-output = 1.1
-cache_read = 0.05
-cache_write = 0
-
-[limit]
-context = 202_752
-output = 202_752
-
-[provider]
-npm = "@ai-sdk/openai-compatible"

providers/crof/models/glm-5.2.toml

@@ -1,16 +1,16 @@
 base_model = "zhipuai/glm-5.2"
-reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high"] }]
 
 [interleaved]
 field = "reasoning_content"
 
-[cost]
-input = 0.3
-output = 1.05
-cache_read = 0.05
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high"]
 
-[limit]
-output = 131_072
+[cost]
+input = 0.15
+output = 0.52
+cache_read = 0.02
 
 [provider]
 npm = "@ai-sdk/openai-compatible"

providers/crof/models/glm-5.toml

@@ -1,20 +0,0 @@
-base_model = "zhipuai/glm-5"
-# Crof's request docs establish no model-specific toggle, effort, or budget.
-# https://crof.ai/docs.md (accessed 2026-06-25)
-reasoning_options = []
-
-[interleaved]
-field = "reasoning_content"
-
-[cost]
-input = 0.48
-output = 1.9
-cache_read = 0.1
-cache_write = 0
-
-[limit]
-context = 202_752
-output = 202_752
-
-[provider]
-npm = "@ai-sdk/openai-compatible"

providers/crof/models/kimi-k2.5-lightning.toml

@@ -1,28 +0,0 @@
-name = "Kimi K2.5 (Lightning)"
-description = "Kimi multimodal agent model for visual understanding, coding, and planning"
-family = "kimi-k2"
-release_date = "2026-02-06"
-last_updated = "2026-02-06"
-attachment = true
-reasoning = true
-reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high"] }]
-temperature = false
-tool_call = true
-structured_output = true
-open_weights = true
-
-[interleaved]
-field = "reasoning_content"
-
-[cost]
-input = 1.00
-output = 3.00
-cache_read = 0.20
-
-[limit]
-context = 131_072
-output = 32_768
-
-[modalities]
-input = ["text", "image", "video"]
-output = ["text"]

providers/crof/models/kimi-k2.5.toml

@@ -1,13 +0,0 @@
-base_model = "moonshotai/kimi-k2.5"
-reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high"] }]
-
-[interleaved]
-field = "reasoning_content"
-
-[cost]
-input = 0.35
-output = 1.7
-cache_read = 0.07
-
-[provider]
-npm = "@ai-sdk/openai-compatible"

providers/crof/models/kimi-k3-eco.toml

@@ -1,10 +1,13 @@
 base_model = "moonshotai/kimi-k3"
 name = "Kimi K3 Eco"
-reasoning_options = [{ type = "effort", values = ["none", "low", "high", "max"] }]
 
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "high", "max"]
+
 [cost]
 input = 1
 output = 4
@@ -12,7 +15,6 @@ cache_read = 0.1
 
 [limit]
 context = 1_000_000
-output = 131_072
 
 [provider]
 npm = "@ai-sdk/openai-compatible"

providers/crof/models/mimo-v2.5-pro.toml

@@ -1,9 +1,12 @@
 base_model = "xiaomi/mimo-v2.5-pro"
-reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high"] }]
 
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high"]
+
 [cost]
 input = 0.4
 output = 0.8
@@ -15,5 +18,8 @@ input = 2
 output = 6
 cache_read = 0.4
 
+[limit]
+context = 1_000_000
+
 [provider]
 npm = "@ai-sdk/openai-compatible"

providers/crof/models/minimax-m2.5.toml

@@ -1,14 +0,0 @@
-base_model = "minimax/MiniMax-M2.5"
-# Crof's request docs establish no model-specific toggle, effort, or budget.
-# https://crof.ai/docs.md (accessed 2026-06-25)
-reasoning = true
-reasoning_options = []
-
-[cost]
-input = 0.11
-output = 0.95
-cache_read = 0.02
-cache_write = 0.375
-
-[provider]
-npm = "@ai-sdk/openai-compatible"

sync.md

@@ -297,6 +297,14 @@ Venice is implemented in `packages/core/src/sync/providers/venice.ts`.
 - Every Venice model uses `base_model`; flattened IDs are matched to provider-agnostic metadata before provider-specific overrides are written.
 - Every Venice model declares `reasoning_options`; models without API-provided effort levels use an empty array.
 
+## CrofAI Notes
+
+- CrofAI is implemented in `packages/core/src/sync/providers/crof.ts` and uses the public `https://crof.ai/v1/models` endpoint.
+- Run it with `bun models:sync crof` or `bun crof:sync`.
+- Pricing, context limits, output limits, and reasoning availability are synchronized from the API for existing models.
+- Local metadata is retained for fields the API does not expose, including descriptions, modalities, model capabilities, and reasoning option values.
+- Models removed from the API are removed locally. New API models are reported but not created automatically because the endpoint does not expose enough metadata for complete catalog entries.
+
 ## Standalone Generators
 
 Some provider scripts in `packages/core/script/generate-*.ts` are not wired into `bun models:sync`. When updating those scripts, preserve existing `base_model` and `base_model_omit` fields for generated TOMLs that already use model metadata inheritance. New inheritance-aware output should use `base_model`; do not reintroduce legacy `[extends]` syntax.