From 9f02148688e81c2cc3f222728cda6e35cc0cde4c Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Tue, 21 Jul 2026 04:04:01 +0000
Subject: [PATCH] Add Kimi (Moonshot AI) provider to Agent Network catalog

Adds a first-class kimi_api catalog entry so operators can govern Kimi K3
and K2 Thinking traffic through the Agent Network proxy instead of riding
the generic custom/vllm entry.

ParserID is left empty so the proxy's URL sniffer dispatches both body
shapes Moonshot serves on the same host and key: the OpenAI-compatible
/v1/chat/completions endpoint and the Anthropic-compatible
/anthropic/v1/messages endpoint that Moonshot's official Claude Code
setup uses (same pattern as the Bifrost gateway entry).

Pricing defaults cover both parser surfaces, including the kimi-k3[1m]
model id some Claude Code guides use for the 1M-context alias, so cost
metering doesn't silently skip those requests.

The dashboard's PROVIDER_CATALOG needs a matching entry (separate repo).
---
 .../modules/agentnetwork/catalog/catalog.go   | 39 +++++++++++++++++++
 .../llm/pricing/defaults_pricing.yaml         | 30 ++++++++++++++
 2 files changed, 69 insertions(+)

diff --git a/management/internals/modules/agentnetwork/catalog/catalog.go b/management/internals/modules/agentnetwork/catalog/catalog.go
index f82cffae6..3c57ce5b7 100644
--- a/management/internals/modules/agentnetwork/catalog/catalog.go
+++ b/management/internals/modules/agentnetwork/catalog/catalog.go
@@ -420,6 +420,45 @@ var providers = []Provider{
 			{ID: "mistral-embed", Label: "Mistral Embed", InputPer1k: 0.0001, OutputPer1k: 0, ContextWindow: 8192},
 		},
 	},
+	{
+		ID:                 "kimi_api",
+		Kind:               KindProvider,
+		Name:               "Kimi (Moonshot AI) API",
+		Description:        "Kimi K3 / K2 models via the Moonshot AI platform",
+		DefaultHost:        "api.moonshot.ai",
+		AuthHeaderName:     "Authorization",
+		AuthHeaderTemplate: "Bearer ${API_KEY}",
+		DefaultContentType: "application/json",
+		BrandColor:         "#1A1A2E",
+		// ParserID empty on purpose: Moonshot serves two body shapes on
+		// the same host and key, and the proxy's URL sniffer dispatches
+		// both (same pattern as Bifrost). /v1/chat/completions matches
+		// OpenAIParser; the Anthropic-compatible endpoint the official
+		// Claude Code guide uses (/anthropic/v1/messages) contains
+		// "/v1/messages" and matches AnthropicParser. Pinning "openai"
+		// here would misparse the Claude Code path — the primary way
+		// teams consume Kimi for coding today. Both endpoints accept the
+		// same Moonshot key via Authorization: Bearer (Claude Code's
+		// ANTHROPIC_AUTH_TOKEN rides that header too).
+		//
+		// api.moonshot.ai is the international platform; mainland-China
+		// accounts live on api.moonshot.cn with separate billing —
+		// operators there override the host on the provider record. The
+		// kimi.com subscription coding endpoint (api.kimi.com/coding,
+		// model id "k3") is account-bound seat licensing rather than a
+		// meterable platform key, so it's deliberately not the default.
+		ParserID: "",
+		// Pricing per Moonshot's platform rates at K3 launch (July 2026):
+		// K3 $3/$15 per MTok with $0.30 cached input, flat across the
+		// 1M-token window; K2 Thinking $0.60/$2.50 with $0.15 cache hits.
+		// Kimi K3 has a single always-on-reasoning SKU — no mini/turbo
+		// variants at launch. The consumer app's "K3 Swarm Max" mode is
+		// not an API SKU, so it doesn't appear here.
+		Models: []Model{
+			{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
+			{ID: "kimi-k2-thinking", Label: "Kimi K2 Thinking", InputPer1k: 0.0006, OutputPer1k: 0.0025, ContextWindow: 262144},
+		},
+	},
 	{
 		ID:                 "litellm_proxy",
 		Kind:               KindGateway,
diff --git a/proxy/internal/llm/pricing/defaults_pricing.yaml b/proxy/internal/llm/pricing/defaults_pricing.yaml
index cd5c64fbf..9f9f4ae78 100644
--- a/proxy/internal/llm/pricing/defaults_pricing.yaml
+++ b/proxy/internal/llm/pricing/defaults_pricing.yaml
@@ -161,6 +161,18 @@ openai:
     input_per_1k: 0.0001
     output_per_1k: 0
 
+  # Kimi / Moonshot AI (kimi_api) — OpenAI-compatible /v1 endpoint. Moonshot
+  # reports cache hits OpenAI-style when present; cached input is 10% of
+  # input for K3 ($0.30 vs $3.00 per MTok) and $0.15/MTok for K2 Thinking.
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cached_input_per_1k: 0.0003
+  kimi-k2-thinking:
+    input_per_1k: 0.0006
+    output_per_1k: 0.0025
+    cached_input_per_1k: 0.00015
+
 anthropic:
   # Claude 4.x family — cache reads ≈10% of input, cache writes ≈125% of input.
   # Pricing source: Anthropic's current published rates per million tokens,
@@ -206,6 +218,24 @@ anthropic:
     cache_read_per_1k: 0.0001
     cache_creation_per_1k: 0.00125
 
+  # Kimi / Moonshot AI (kimi_api) via the Anthropic-compatible endpoint
+  # (/anthropic/v1/messages — the official Claude Code setup). Same rates
+  # as the OpenAI-shape entries above. "kimi-k3[1m]" is the model id some
+  # Claude Code guides set for the 1M-context alias; priced identically so
+  # cost metering doesn't silently skip those requests.
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+  "kimi-k3[1m]":
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+  kimi-k2-thinking:
+    input_per_1k: 0.0006
+    output_per_1k: 0.0025
+    cache_read_per_1k: 0.00015
+
 bedrock:
   # AWS Bedrock model ids, normalised by the request parser (cross-region
   # inference-profile prefix + version/throughput suffix stripped), e.g.