Add Vercel AI Gateway as a first-class provider

README.md

@@ -54,7 +54,7 @@ Under the hood:
 
 Model access is yours: pick a provider, paste your key, switch anytime. Supported out of the box:
 
-**OpenAI · Anthropic · Google Gemini · Inkling (Thinking Machines) · GLM (Z.ai) · DeepSeek · Kimi (Moonshot) · Qwen · MiniMax · Mistral · Grok (xAI)** - plus open-weight models via **Together** and **Fireworks**, and fully local models via **Ollama**.
+**OpenAI · Anthropic · Google Gemini · Inkling (Thinking Machines) · GLM (Z.ai) · DeepSeek · Kimi (Moonshot) · Qwen · MiniMax · Mistral · Grok (xAI)** - plus open-weight models via **Together** and **Fireworks**, hundreds of models behind one key via the **Vercel AI Gateway**, and fully local models via **Ollama**.
 
 A curated model list marks what we've verified for tool-calling work. Adding any model string works at your own risk.
 

coworker/providers/matrix.py

@@ -16,9 +16,9 @@
 showing a made-up denominator. Values entered 2026-07-28 from vendor docs; verify alongside
 the id refresh.
 
-Resellers: Together + Fireworks + OpenRouter. TODO: add Groq entries here AND its
-descriptor in ``registry.py`` once the current provider surface is tested — deliberately
-deferred to bound how much needs verifying at once.
+Resellers: Together + Fireworks + OpenRouter + Vercel AI Gateway. TODO: add Groq
+entries here AND its descriptor in ``registry.py`` once the current provider surface
+is tested — deliberately deferred to bound how much needs verifying at once.
 """
 
 from __future__ import annotations
@@ -31,12 +31,17 @@
 _AGENTIC = ModelCapabilities(
     tools=True, vision=False, parallel_tool_calls=True, streaming=True
 )
-# The native three (OpenAI, Anthropic, Gemini) all take PDFs directly; every
-# OpenAI-compatible vendor and reseller in the matrix does not (their chat APIs have
-# no inline file part — checked 2026-07-17), so those fall back via pdf_support.py.
+# The native three (OpenAI, Anthropic, Gemini) all take PDFs directly; compat vendors
+# and resellers fall back via pdf_support.py instead (no inline file part — checked
+# 2026-07-17; the Vercel gateway documents one, but its shape is unverified here, so
+# it stays on the fallback too).
 _AGENTIC_VISION = ModelCapabilities(
     tools=True, vision=True, pdf=True, parallel_tool_calls=True, streaming=True
 )
+# Vision over an OpenAI-compat surface: tools and images, PDFs on the fallback (above).
+_AGENTIC_VISION_COMPAT = ModelCapabilities(
+    tools=True, vision=True, parallel_tool_calls=True, streaming=True
+)
 
 
 @dataclass(frozen=True)
@@ -89,12 +94,7 @@ class ModelEntry:
     # Muse Spark (Meta Model API, public preview 2026-07-09): multimodal + tools via
     # their OpenAI-compat surface. Vision yes; PDFs unverified over compat — falls
     # back via pdf_support.py like the other compat vendors.
-    "meta:muse-spark-1.1": ModelEntry(
-        "Muse Spark 1.1 · Meta",
-        ModelCapabilities(
-            tools=True, vision=True, parallel_tool_calls=True, streaming=True
-        ),
-    ),
+    "meta:muse-spark-1.1": ModelEntry("Muse Spark 1.1 · Meta", _AGENTIC_VISION_COMPAT),
     "zai:glm-5.2": ModelEntry("GLM-5.2 · Z AI", _AGENTIC, 128_000),
     "deepseek:deepseek-v4-flash": ModelEntry(
         "DeepSeek V4 Flash · DeepSeek", _AGENTIC, 128_000
@@ -115,11 +115,7 @@ class ModelEntry:
     # Kimi K3 on Together (landed late July 2026): 1M window, native vision; PDFs
     # unverified over the compat surface (falls back via pdf_support.py, like Muse Spark).
     "together:moonshotai/Kimi-K3": ModelEntry(
-        "Kimi K3 · via Together",
-        ModelCapabilities(
-            tools=True, vision=True, parallel_tool_calls=True, streaming=True
-        ),
-        1_000_000,
+        "Kimi K3 · via Together", _AGENTIC_VISION_COMPAT, 1_000_000
     ),
     "together:moonshotai/Kimi-K2.7-Code": ModelEntry(
         "Kimi K2.7 Code · via Together", _AGENTIC, 256_000
@@ -157,6 +153,29 @@ class ModelEntry:
     "openrouter:meta-llama/llama-4-maverick": ModelEntry(
         "Llama 4 Maverick · via OpenRouter", _AGENTIC, 1_000_000
     ),
+    # Vercel AI Gateway: creator/model ids, verbatim. Curated to agent-capable,
+    # cost-efficient tiers, one per major creator. Windows and vision are the
+    # gateway's own /v1/models values (checked 2026-08-05) and are per-route —
+    # GLM-5.2 is 1M here vs 128k via Together. The catalog's pdf modality is
+    # deliberately unmapped (_AGENTIC_VISION_COMPAT: file-part shape unverified).
+    "vercel:anthropic/claude-haiku-4.5": ModelEntry(
+        "Claude Haiku 4.5 · via Vercel AI Gateway", _AGENTIC_VISION_COMPAT, 200_000
+    ),
+    "vercel:zai/glm-5.2": ModelEntry(
+        "GLM-5.2 · via Vercel AI Gateway", _AGENTIC, 1_000_000
+    ),
+    "vercel:deepseek/deepseek-v4-flash": ModelEntry(
+        "DeepSeek V4 Flash · via Vercel AI Gateway", _AGENTIC, 1_000_000
+    ),
+    "vercel:openai/gpt-5.4-mini": ModelEntry(
+        "GPT-5.4 mini · via Vercel AI Gateway", _AGENTIC_VISION_COMPAT, 400_000
+    ),
+    "vercel:google/gemini-3.6-flash": ModelEntry(
+        "Gemini 3.6 Flash · via Vercel AI Gateway", _AGENTIC_VISION_COMPAT, 1_000_000
+    ),
+    "vercel:moonshotai/kimi-k2.7-code": ModelEntry(
+        "Kimi K2.7 Code · via Vercel AI Gateway", _AGENTIC_VISION_COMPAT, 256_000
+    ),
     # -- cloud accounts (models running in the user's own AWS/GCP) ----------------
     # Bedrock ids carry a family segment (claude/ → native Anthropic path, other/ →
     # Converse) plus AWS's own `-v<n>:<m>` version suffix. Some regions require the

coworker/providers/registry.py

@@ -550,6 +550,13 @@ def _compat(
         recommended_model="z-ai/glm-5.2",
         env_key="OPENROUTER_API_KEY",
     ),
+    _compat(
+        "vercel",
+        "Vercel AI Gateway",
+        base_url="https://ai-gateway.vercel.sh/v1",
+        recommended_model="anthropic/claude-haiku-4.5",
+        env_key="AI_GATEWAY_API_KEY",  # Vercel's documented name for this key
+    ),
     ProviderDescriptor(
         name="ollama",
         title="Ollama (local models)",

surfaces/gui/src/providers/ProviderSetup.tsx

@@ -25,6 +25,11 @@ export const KEY_HELP: Record<string, { url: string; label: string }> = {
   bedrock: { url: "https://console.aws.amazon.com/bedrock/home#/api-keys", label: "the AWS Bedrock console" },
   fireworks: { url: "https://fireworks.ai/account/api-keys", label: "fireworks.ai" },
   together: { url: "https://api.together.xyz/settings/api-keys", label: "together.xyz" },
+  // Vercel's own docs deep-link (resolves the signed-in team's dashboard).
+  vercel: {
+    url: "https://vercel.com/d?to=%2F%5Bteam%5D%2F%7E%2Fai-gateway%2Fapi-keys&title=AI+Gateway+API+Keys",
+    label: "vercel.com",
+  },
   zai: { url: "https://z.ai/manage-apikey/apikey-list", label: "z.ai" },
   kimi: { url: "https://platform.moonshot.ai/console/api-keys", label: "platform.moonshot.ai" },
   deepseek: { url: "https://platform.deepseek.com/api_keys", label: "platform.deepseek.com" },

surfaces/gui/src/providers/logos.ts

@@ -14,6 +14,7 @@ import vertex from "./logos/vertex.svg";
 import openrouter from "./logos/openrouter.svg";
 import fireworks from "./logos/fireworks.svg";
 import together from "./logos/together.svg";
+import vercel from "./logos/vercel.svg";
 import zai from "./logos/zai.svg";
 import kimi from "./logos/kimi.svg";
 import deepseek from "./logos/deepseek.svg";
@@ -34,6 +35,7 @@ export const PROVIDER_LOGOS: Record<string, string> = {
   openrouter,
   fireworks,
   together,
+  vercel,
   zai,
   kimi,
   deepseek,
@@ -54,6 +56,7 @@ export const PROVIDER_ORDER = [
   "openrouter",
   "fireworks",
   "together",
+  "vercel",
   "zai",
   "kimi",
   "deepseek",

tests/test_providers.py

@@ -352,6 +352,7 @@ def test_compat_models_route_and_get_tool_capabilities():
         "qwen:qwen3-max",
         "xai:grok-4.3",
         "mistral:mistral-large-latest",
+        "vercel:zai/glm-5.2",
     ):
         prefix = model.split(":", 1)[0]
         assert router._provider_name(model) == prefix
@@ -383,6 +384,8 @@ def test_matrix_answers_capabilities_for_reseller_ids():
         "fireworks:accounts/fireworks/models/kimi-k2p6",
         "openrouter:z-ai/glm-5.2",
         "openrouter:meta-llama/llama-4-maverick",
+        "vercel:anthropic/claude-haiku-4.5",
+        "vercel:moonshotai/kimi-k2.7-code",
     ):
         caps = capabilities_for(mid)
         assert caps.tools and caps.parallel_tool_calls and caps.streaming
@@ -409,7 +412,7 @@ def test_reseller_descriptors_and_matrix_stay_in_lockstep():
     from coworker.providers.matrix import models_for_provider
     from coworker.providers.registry import get_descriptor
 
-    for name in ("together", "fireworks", "openrouter"):
+    for name in ("together", "fireworks", "openrouter", "vercel"):
         d = get_descriptor(name)
         assert d is not None and d.needs_key
         curated = models_for_provider(name)