Moonshot AI / Kimi — Specialty LLM Tier

One provider, two product names: Moonshot AI is the company; Kimi is the consumer product and model family. In OpenClaw the two names are used interchangeably. The active integration is the Kimi K2 API (OpenAI-compatible endpoint at https://api.moonshot.ai/v1) accessed via portkey virtual key moonshot-kimi-84f4a1. This is a Tier 2 specialty path used for reasoning-heavy or coding-adjacent workloads — not the primary LLM tier. A SOURCE MISSING flag applies: the KB at knowledge-base/moonshot/ contains only a vendor audit file (_audit/2026-05-02.md), not a formal API.md. All documentation in this hub is sourced from that audit file + proxy.js code inspection.

⚠️ SOURCE MISSING

knowledge-base/moonshot/ contains only _audit/2026-05-02.md — no API.md, no WEBHOOKS.md, no CREDENTIALS.md. All content in this hub is derived from:

workspace/knowledge-base/moonshot/_audit/2026-05-02.md (vendor audit, 39 findings, 2026-05-02)
/home/opsadmin/.openclaw/portkey-proxy/proxy.js — Fix #1 implementation (lines 344-360)
workspace/_hubs/integrations/portkey.md — VK moonshot-kimi-84f4a1 entry

G-KB-SYNC-WITH-CLAUDEMD note: moonshot is in CLAUDE.md “Platforms with KB docs” list, but only has an _audit subdirectory, not a full API.md. Recommend creating knowledge-base/moonshot/API.md from vendor docs at https://platform.kimi.ai/docs/guide/start-using-kimi-api and https://github.com/moonshotai/Kimi-K2.

Quick reference

Field	Value
Vendor	Moonshot AI (product: Kimi)
URL	https://platform.moonshot.ai / https://platform.kimi.ai
KB doc	`[[sources/kb/moonshot/_audit/2026-05-02]]` (audit only — SOURCE MISSING for API.md)
Auth method	Bearer (API key)
Auth credential	`op://Aurora/moonshot/api_key`
Cred-proxy port	n/a (pending B1-B6 ratification per nemoclaw-audit-2026-05-03)
Webhook port	n/a
Webhook handler	n/a
Webhook dedup table	n/a
Tunnel path	n/a
Outbound API base	`https://api.moonshot.ai/v1` (OpenAI-compatible)
Anthropic-compat base	`https://api.moonshot.ai/anthropic` (Anthropic-format; NOT used in production)
Rate limits	RPM scales with cumulative recharge tier (default: low; paid: hundreds/min)
Rate-limit action	429 → exponential backoff (3 retries), Discord ops alert; 429 may mask balance exhaustion
Cost	kimi-k2-turbo: up to 50% lower input pricing; kimi-k2-thinking-turbo: same reduced rates
Backup/recovery	Vendor-owned; fall through to OpenRouter via Portkey tier config on failure
Discord alert channel	ops
Drift cadence	manual
Status	production
Primary models	`kimi-k2.6`, `kimi-k2-turbo`, `kimi-k2-thinking`, `kimi-k2-thinking-turbo`
Portkey virtual key	`moonshot-kimi-84f4a1`

Fix #1: Kimi 404 → per-model gating in proxy.js

This is the most important implementation detail for this integration. Without this fix, Kimi (or any non-Claude model) requested by a max-plan agent caused a 404 on the :18903 LOCAL-SHORTCUT path, because that upstream binary (anthropic-max-router) only handles Anthropic /v1/messages requests.

Root cause

Max-plan agents (served via :18903) have their traffic routed to the LOCAL-SHORTCUT path by default. The :18903 anthropic-max-router does not serve non-Anthropic models — it returns 404 on anything that isn’t a Claude model request.

Fix implementation

proxy.js lines 344-360 gate the LOCAL-SHORTCUT path on isAnthropicModel:

// Step 2: Determine tier now that model is known.
// Max-plan agent routing is gated by isAnthropicModel — non-Claude models on max-plan
// agents (e.g., dispo asking for kimi-for-coding) fall through to Portkey cloud path.
// Without this gate, ALL of the agent's traffic would hit the Anthropic-only LOCAL-SHORTCUT
// and 404 on non-Claude models.
const isAnthropicModel = !!(model && model.toLowerCase().startsWith('claude'));
let portKeyConfig, tier;
const isMaxPlanAgent = !!(agentId && AGENT_CONFIGS[agentId] && isAnthropicModel);
if (isMaxPlanAgent) {
  const agentCfg = AGENT_CONFIGS[agentId];
  portKeyConfig = agentCfg.config;
  tier = agentCfg.tier;
} else {
  const tierInfo = MODEL_CONFIGS[model] || { config: DEFAULT_CONFIG, tier: 'default' };
  portKeyConfig = tierInfo.config;
  tier = tierInfo.tier;
}

Effect: When a max-plan agent requests kimi-k2.6 (or any non-claude- model):

isAnthropicModel = false
isMaxPlanAgent = false (gated)
Falls through to MODEL_CONFIGS[model] → moonshot-kimi-84f4a1 virtual key → Portkey cloud → Moonshot API
No 404. Tool_calls row written via CHOKEPOINT-1.

Fix location: /home/opsadmin/.openclaw/portkey-proxy/proxy.js, lines ~344–360. Fix authored: 2026-05-01 (per portkey hub notes as “Fix #1”). Also documented in: portkey and anthropic.

Kimi K2 model family

Model	Best for	Notes (from audit)
`kimi-k2.6`	General-purpose Kimi	OpenAI SDK compatible; `base_url=https://api.moonshot.ai/v1`
`kimi-k2-turbo`	High-speed, input-heavy	50% lower input pricing; ~100 tokens/second throughput
`kimi-k2-thinking`	Complex multi-step reasoning	Tool use, function calling, agentic tasks
`kimi-k2-thinking-turbo`	Reasoning + speed	Same reduced pricing as turbo; speed + reasoning combined
`kimi-k2-0905-preview`	Preview / testing	Earlier preview; not for production

Key capability from audit: Kimi K2 is a 1T parameter MoE model with 32B activated parameters, specifically designed for tool use, reasoning, and autonomous problem-solving. Strong on tau2-telecom and AceBench benchmarks — well-suited for multi-step RE workflow orchestration.

Temperature note: Kimi-K2-Instruct recommended temperature = 0.6. If using Anthropic-compatible endpoint, Moonshot applies implicit mapping real_temperature = request_temperature * 0.6 — do not rely on this; use OpenAI-compatible endpoint with direct temperature=0.6.

Components

/home/opsadmin/.openclaw/portkey-proxy/proxy.js — Fix #1 implementation (lines 344–360); isAnthropicModel gate; moonshot-kimi-84f4a1 VK assignment via MODEL_CONFIGS
workspace/knowledge-base/moonshot/_audit/2026-05-02.md — vendor audit (39 findings); SOURCE MISSING for formal API reference
Portkey virtual key moonshot-kimi-84f4a1 — resolves to op://Aurora/moonshot/api_key in Portkey dashboard
Platform dashboard: https://platform.moonshot.ai — API key management, usage, billing recharge tiers

How it’s used

Trigger condition: Agent requests a model starting with kimi- (e.g., kimi-k2.6, kimi-k2-turbo) or moonshot/ prefixed model; or a Portkey tier config specifies Moonshot for reasoning tasks
Workflow: proxy.js → isAnthropicModel = false (kimi does not start with ‘claude’) → MODEL_CONFIGS["kimi-k2.6"] → Portkey VK moonshot-kimi-84f4a1 → https://api.moonshot.ai/v1/chat/completions → Kimi response
Agents involved: dispo (kimi-for-coding pattern noted in portkey hub); any agent needing agentic/reasoning tasks that benefit from K2’s tool-call strength
Failure mode: 404 (prevented by Fix #1); 429 rate limit (RPM tied to billing tier — may need recharge); balance exhaustion returns 429 (indistinguishable from rate limit)
Success criteria: Non-404 response; tool_calls row written to Supabase; latency appropriate for model variant (turbo = faster)

Cross-links

Agents that touch this

_summary — “kimi-for-coding” model request pattern documented in portkey hub
_summary — may route reasoning-heavy sub-tasks to Kimi K2
_summary — complex multi-step research; K2 agentic tool-calling is relevant

Skills that invoke this

acq-analyze — complex analysis tasks may benefit from K2 reasoning
hubspot-deal-ingest — multi-hop data enrichment; K2 tool-calling could orchestrate

Plans that govern this

openclaw-self-improvement-layer-2026-05-03 — OSIL model tier evaluation; Kimi K2 is Tier 2 candidate for agentic tasks
vendor-deep-audit-comprehensive-2026-05-02 — Moonshot/Kimi in Tier 1 vendor deep-audit candidates
nemoclaw-audit-2026-05-03 — B1-B6 ratification; cred-proxy eligibility for Moonshot endpoint
openclaw-fragmentation-fix-2026-05-01 — G-NO-PLAINTEXT-CREDS; key must be op://Aurora/moonshot/api_key

Feedback rules

feedback_no_plaintext_creds — op://Aurora/moonshot/api_key only; never inline
feedback_chokepoint_principle — G-CHOKEPOINT-1; all Kimi calls must write tool_calls row
feedback_live_over_memory — SOURCE MISSING: verify actual Moonshot API behavior from live docs, not audit inference
feedback_verify_schema_before_designing — check actual Kimi response shape before writing parsing code; audit is low-confidence
feedback_audit_before_architect — run a live Kimi test call before building Moonshot-dependent features
feedback_dual_write_required — G-DUAL-WRITE; Moonshot responses must reach both Portkey observability and Supabase

KB / source docs

[[sources/kb/moonshot/_audit/2026-05-02]] — 39-finding vendor audit (confidence: low throughout; NEEDS VERIFICATION for production use)
SOURCE MISSING: No knowledge-base/moonshot/API.md — recommended remediation: create from https://platform.kimi.ai/docs/guide/start-using-kimi-api

System maps

request-lifecycle — non-Anthropic model path; same flow as OpenRouter but terminates at Moonshot
ports-topology — Portkey :18900 as single routing surface

Anchor: portkey

Member	Role
portkey	Anchor; CHOKEPOINT-1; Fix #1 lives in proxy.js; VK `moonshot-kimi-84f4a1`
anthropic	Primary tier (Tier 1); Moonshot is Tier 2 specialty alternative
openrouter	Tier 2 aggregator; broader model selection; different VK (`openrouter-fallback`)
moonshot-kimi	This hub — Tier 2 specialty; Kimi K2 MoE; tool-use + reasoning
voyage	Embedding tier; separate from LLM routing; Memory/embeddings cluster anchor: supabase

Open issues / TODOs

SOURCE MISSING: Create knowledge-base/moonshot/API.md from https://platform.kimi.ai/docs/guide/start-using-kimi-api and https://github.com/moonshotai/Kimi-K2
cred-proxy-eligible: pending — resolve after B1-B6 ratification per nemoclaw-audit-2026-05-03
Verify which MODEL_CONFIGS keys in proxy.js map to moonshot-kimi-84f4a1 — may need kimi-k2.6, kimi-k2-turbo, kimi-k2-thinking all registered
Billing tier: Moonshot RPM scales with cumulative recharge — verify current tier before enabling high-volume routing
Anthropic-compatible endpoint (https://api.moonshot.ai/anthropic) is NOT wired and NOT tested — do not use without explicit testing
Temperature behavior: audit notes implicit temperature * 0.6 mapping on Anthropic-compat path — OpenAI-compat path should use direct temperature=0.6 per Kimi docs

Recent activity

2026-05-03: hub created (W2-S7); SOURCE MISSING flagged; Fix #1 documented
2026-05-01: Fix #1 (Kimi 404 → per-model gating) shipped in proxy.js
2026-05-02: Vendor audit knowledge-base/moonshot/_audit/2026-05-02.md generated (39 findings)

Quartz 4

Explorer

Moonshot AI / Kimi — Specialty LLM Tier

Moonshot AI / Kimi — Specialty LLM Tier

⚠️ SOURCE MISSING

Quick reference

Fix #1: Kimi 404 → per-model gating in proxy.js

Root cause

Fix implementation

Kimi K2 model family

Components

How it’s used

Cross-links

Agents that touch this

Skills that invoke this

Plans that govern this

Feedback rules

KB / source docs

System maps

Open issues / TODOs

Recent activity

Graph View

Table of Contents

Backlinks

Quartz 4

Explorer

Moonshot AI / Kimi — Specialty LLM Tier

Moonshot AI / Kimi — Specialty LLM Tier

⚠️ SOURCE MISSING

Quick reference

Fix #1: Kimi 404 → per-model gating in proxy.js

Root cause

Fix implementation

Kimi K2 model family

Components

How it’s used

Cross-links

Agents that touch this

Skills that invoke this

Plans that govern this

Feedback rules

KB / source docs

System maps

Related: LLM provider tier cluster

Open issues / TODOs

Recent activity

Graph View

Table of Contents

Backlinks