Kimi K3 runs in almost every serious CLI and terminal coding tool today: Codex CLI, Claude Code, OpenCode, pi, Crush, Goose, Hermes Agent. All of them accept a custom OpenAI-compatible (or Anthropic-compatible) endpoint, and that is all a K3 connection takes: a base URL, an API key, and the model ID kimi-k3. This page is the hub: one table that sends you to the verified walkthrough for your tool.
One boundary, up front. This page answers which tool to put Kimi K3 in and how to configure it. If your question is where to get access in the first place (official API, per-token providers, gateways, subscriptions, self-hosting), that is a different decision, covered in Where to get Kimi K3: every access route compared. The two pages meet at the API key: get one there, spend it here. Building in TypeScript rather than picking a tool? That is the SDK side of the same boundary: Kimi K3 in the Vercel AI SDK, one npm package and two lines of code.
Everything below is as of October 11, 2026. Disclosure: Wallaby Token sells API access to Kimi K3 ($2.70 input / $13.50 output per 1M tokens, 10% off the official list, ongoing), and the endpoint in our guides is ours.
The tool table
| Tool | Type | Setup in one line | Walkthrough | Verified with |
|---|---|---|---|---|
| Codex CLI | Terminal agent | Profile TOML with wire_api = "responses"; Responses API only, "chat" now hard-errors |
Kimi K3 in Codex CLI | Codex CLI 0.154.0 (tested 2026-09-13) |
| Claude Code | Terminal agent | Three env vars (ANTHROPIC_BASE_URL, token, model) on an Anthropic-compatible endpoint; no /v1 suffix |
Kimi K3 in Claude Code | Claude Code (tested 2026-09-09) |
| OpenCode | Terminal TUI | Zero config: Wallaby is in the models.dev catalog, so one env var plus opencode models --refresh |
Kimi K3 in OpenCode | opencode 1.18.31 (tested 2026-09-17) |
| pi | Terminal agent | pi install npm:pi-wallaby-provider (npm) on pi 1.0+, or one JSON block in models.json on 0.86.x |
Three commands, no config files · One JSON block | pi 1.0.4 / 0.86.1 (tested 2026-10-11) |
| Crush | Terminal TUI | crushrc with an openai-compat provider, plus price fields so the TUI shows real session cost |
Kimi K3 in Crush | crush 0.98.1 (tested 2026-10-10) |
| Goose | Terminal CLI | Two env vars (OPENAI_HOST must omit /v1) plus two YAML lines |
Kimi K3 in Goose | goose 1.54.0 (tested 2026-10-10) |
| Hermes Agent | Terminal CLI | Four lines of YAML in ~/.hermes/config.yaml; /v1 suffix required |
Kimi K3 in Hermes Agent | Hermes 0.19.0 (tested 2026-09-10) |
| Cline | VS Code extension | Built-in Wallaby provider since v4.1.19: select Wallaby, paste base URL and key, pick kimi-k3 |
Kimi K3 in Cline | Cline 4.1.22 (tested 2026-10-07) |
| VS Code Copilot | Editor (BYOK) | Custom Endpoint provider, chat-completions API type; covers Chat and agent mode |
Kimi K3 in VS Code via BYOK | VS Code BYOK (tested 2026-09-12) |
| TRAE | IDE (BYOK) | Custom Model with OpenAI Chat Completions format; ignore the hardwired "OpenAI" entry | TRAE BYOK, tested | TraeCode 3.6.2 (tested 2026-10-10) |
| OpenHands | Web agent UI | One saved LLM profile: openai/kimi-k3, base URL, key |
OpenHands setup and first-chat fix | agent-server 1.36.0 (tested 2026-09-23) |
| Cursor | Editor | As of Oct 11, 2026: does not work with third-party endpoints — K3 exists only as Cursor's built-in, paywalled offering on free plans | Why Kimi K3 is blocked in Cursor | Cursor (tested 2026-10-07) |
The shared mental model
Almost every row above reduces to the same three values: base URL (https://api.wallabytoken.com/v1 for our endpoint), API key, and model ID (kimi-k3 exactly, not the namespaced moonshotai/kimi-k3). Two recurring traps:
- The
/v1suffix is per-tool. Goose and Claude Code append their own path and need the bare host; Hermes and most OpenAI-compatible clients need the full/v1. A 404 on every request almost always means this. - Codex speaks the Responses API only. Current Codex versions removed chat-completions support entirely, so the provider block must set
wire_api = "responses"against an endpoint that accepts Responses traffic natively (current Codex CLI and desktop releases; older builds are unmaintained). Any older guide telling you otherwise is stale.
For agent features like file edits and shell commands, the model must support tool calling. K3 does.
Your tool is not listed?
If a tool can point at a custom OpenAI-compatible endpoint, it can run Kimi K3. Find its "custom provider" or "OpenAI compatible" setting, enter the three values above, then verify with a small real task: one that exercises tool use, not just chat. Check the provider's docs for whether it appends /v1 itself, and confirm the model name string exactly.
Try it for $0.50
New Wallaby accounts get $0.50 in trial credit, no card: enough for dozens of verification runs across several tools in this table. Pick your row, follow the walkthrough, and watch the itemized cost of your first real task in the console. If you are still deciding on the access route itself, start with the route comparison and come back here with a key.
Note: configuration options, version numbers, and pricing in this post were verified on October 11, 2026. Tools and APIs change — cross-check the official documentation before relying on specifics.