Kimi K3 in OpenClaw

OpenClaw's onboarding default is moonshot/kimi-k3 on Moonshot's own API at $3 / $15. The same model on ainetcafe is $2.10 / $10.50 with $0.30 cached input, and OpenClaw's memory and system prompt hit the cache on every turn. You do not need a plugin: OpenClaw lets you register any OpenAI-compatible endpoint as a custom provider.

Option A: plugin (one command)

openclaw plugins install npm:openclaw-ainetcafe-provider --force --accept-capabilities
export AINETCAFE_API_KEY="sk-..."
openclaw models list --provider ainetcafe --refresh
openclaw agent --local --model ainetcafe/Kimi-K3 --thinking low -m "Say OK"
openclaw gateway restart   # if a gateway is already running

--force is needed because npm is outside ClawHub review; the plugin's source is public on GitHub (one 70-line file plus a manifest). It registers provider ainetcafe with the same model definition as Option B, adds the --ainetcafe-api-key onboarding flag, and was tested on OpenClaw 2026.9.4 on 17 September: a plain turn and an exec tool-call turn both completed with winnerProvider: ainetcafe. Prefer Option B if you do not want a plugin at all.

Option B: config only

1. Export the key

export AINETCAFE_API_KEY="sk-..."   # add to your shell profile, or put it under env.vars in the config

2. Add the provider to ~/.openclaw/openclaw.json

Merge this into your existing config (JSON5, comments allowed). If the file does not exist yet, this is a complete one.

{
  agents: {
    defaults: {
      model: { primary: "ainetcafe/Kimi-K3" },
      models: { "ainetcafe/Kimi-K3": { alias: "Kimi K3 (ainetcafe)" } },
    },
  },
  models: {
    mode: "merge",
    providers: {
      ainetcafe: {
        baseUrl: "https://microquickjs.com/v1",
        apiKey: "${AINETCAFE_API_KEY}",
        api: "openai-completions",
        models: [
          {
            id: "Kimi-K3",
            name: "Kimi K3 (ainetcafe)",
            reasoning: true,
            input: ["text", "image"],
            cost: { input: 2.1, output: 10.5, cacheRead: 0.3, cacheWrite: 0 },
            contextWindow: 262144,
            maxTokens: 65536,
            compat: {
              supportsReasoningEffort: true,
              supportedReasoningEfforts: ["low", "high", "max"],
              supportsStore: false,
              supportsDeveloperRole: false,
              supportsStrictMode: false,
              maxTokensField: "max_tokens",
              requiresReasoningContentOnAssistantMessages: true,
            },
          },
        ],
      },
    },
  },
}

What the fields do: reasoning: true lets /think and --thinking pick a level; supportedReasoningEfforts maps those to K3's low / high / max; requiresReasoningContentOnAssistantMessages keeps K3's reasoning_content in the replayed history, which K3 needs for multi-step tool loops; cost makes OpenClaw's spend estimates right. contextWindow is 256K on the standard plan; set it to 1048576 if we have enabled 1M context for your account.

3. Check and run

openclaw models list --provider ainetcafe
# ainetcafe/Kimi-K3   text+image   262k   default,configured

openclaw agent --local --thinking low -m "Say OK"
# OK

openclaw gateway restart   # if the gateway is already running, so it picks up the new provider

This is exactly what we ran on 17 September against OpenClaw 2026.9.4: the plain turn returned in one attempt (winnerProvider: ainetcafe, fallbackUsed: false), and a turn that required the exec tool made one tool call with zero failures.

Non-interactive setup

For scripts and CI, OpenClaw's onboarding accepts a custom base URL directly:

openclaw onboard --non-interactive --accept-risk --skip-health \
  --mode local \
  --custom-base-url "https://microquickjs.com/v1" \
  --custom-api-key "$AINETCAFE_API_KEY" \
  --custom-model-id "Kimi-K3"

That path registers a generic custom provider without the compat block above; for long tool loops, prefer the explicit config.

Why K3 fits OpenClaw

• OpenClaw ships forty-plus tools and long-running sessions; K3 keeps structured tool calls stable across long loops (402 of 408 schema cases in Moonshot's verifier, streaming and non-streaming).
• Its system prompt and memory are re-sent every turn. With prompt caching that context bills at $0.30 per million instead of $2.10.
• Vision is on: browser screenshots and image attachments go straight to the model (input: ["text", "image"]).

Verify the key first

Thirty seconds now saves an hour of guessing later. Run this in a terminal; you should see a short reply and a usage block.

curl https://microquickjs.com/v1/chat/completions \
  -H "Authorization: Bearer $AINETCAFE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"Kimi-K3","messages":[{"role":"user","content":"Say OK"}],"max_tokens":5}'

If something goes wrong

Still stuck? Email [email protected] with the request ID from the error body.