Moonshot Kimi

Moonshot's Kimi Coding 2.6 (released April 2026) is available as a Pi-SDK-backed provider in Fabric Agents. Access is via API key.

Get an API key

1. Sign in at platform.kimi.com (or platform.moonshot.cn if you're on the Chinese mainland instance).
2. Create a key under API Keys — it starts with sk-kimi-.
3. Top up a balance or confirm you're on Moonshot's free trial — keys with zero quota silently return "quota exceeded" on first send.

Connect in Fabric Agents

1. Open Settings → AI → Connections → Add.
2. From the provider picker, choose Kimi (Coding) — this maps to endpoint https://api.kimi.com/coding.
3. Paste your sk-kimi-… key.
4. Click Test connection. You should see the model catalog populate with Kimi Coding 2.6 and Kimi K2 Thinking.
5. Save.

The connection shows up in the model picker. Sessions created with this connection are routed through the Pi SDK using Moonshot's coding-optimized inference backend.

Model ids

Moonshot moved off version-stamped ids for this product line in April 2026 — the API now exposes a single endpoint-name id that always points at the latest coding model.

Fabric Agents auto-migrates older configs — if you set up Kimi before April 2026 with k2p5 in your model list, the next launch replaces it with kimi-for-coding and marks that as the Best-tier default.

Model capabilities

Both Kimi models under this provider:

• Use the anthropic-messages API family, so tool calling and streaming behave the same as Claude connections.
• Accept text and image inputs (kimi-for-coding) or text only (kimi-k2-thinking).
• Support a 256 K token context window with a 32 K max-output per turn.
• Support extended reasoning / thinking. In Fabric Agents, the thinking-level selector in the input bar maps to Moonshot's reasoning_effort.

Tier picker defaults

When you connect Kimi for the first time, the onboarding 3-tier defaults resolve to:

You can override any of them in Settings → AI → Kimi (Coding) → Models. Sessions then use the Balanced tier by default; pick Best or Fast from the model picker per session.

Regional notes

• Moonshot runs two separate clouds — global (platform.kimi.com, endpoint api.kimi.com) and China (platform.moonshot.cn, endpoint api.moonshot.cn). Your sk-kimi-… key is tied to one of them; they don't cross-authenticate. If you've got the wrong region on file, Test connection returns 401.
• The China cloud's rate limits are stricter; for heavy automated use, Moonshot's docs recommend a dedicated enterprise key.

Troubleshooting

Settings shows "Kimi K2.5" instead of 2.6 — you're on a build prior to v0.8.12. Upgrade the desktop app (the in-app auto-updater picks up stable releases from agents.fabric.pro/electron/latest/). The fix adds the kimi-for-coding id to the picker and migrates older k2p5 saves.

Test connection returns 401 — wrong cloud (see Regional notes) or the key was revoked. Cycle it on the Moonshot dashboard and paste the new one.

"model not found" when sending a message — your connection's model list still has k2p5 but Moonshot's API no longer accepts it. Open Settings → AI → Kimi (Coding) → Models, remove k2p5, add kimi-for-coding, save. (The auto-migration in v0.8.12+ handles this on launch; this is only a manual step if you skipped the upgrade.)

Long tool-using sessions degrade / lose context — bump the thinking level for that session. Kimi's reasoning_effort tiers substantially affect chain-of-thought depth on complex agentic tasks.

Related

• LLM Providers overview — other provider setups
• Interactions reference — thinking levels, model picker behaviour
• Release notes for v0.8.12 — the version that added Kimi Coding 2.6