Custom endpoints (GLM, Qwen, Kimi, DeepSeek, MiniMax)

Many model vendors serve the Anthropic Messages API directly: GLM on Z.AI, Qwen on Alibaba Model Studio, Kimi on Moonshot, DeepSeek, MiniMax, and others. Kanna can run them through the same Claude agent it uses for Claude itself, so tools, skills, MCP servers, plan mode and subagents all work unchanged.

Only Anthropic-compatible endpoints work here. A vendor that offers only an OpenAI-compatible API cannot be added.

Add an endpoint

1. Open Settings → Providers and find Custom endpoints.
2. Click Add endpoint and pick a preset (GLM, Qwen, Kimi, DeepSeek, MiniMax, Ollama), or Custom… for any other URL. A preset fills in the base URL and a few model ids; every field stays editable.
3. Paste your API key from the vendor’s console.
4. Check the models list against the vendor’s current model ids. If you know the model’s real Context window, enter it in tokens (optional). Then click Add endpoint.

The id (for example glm) becomes part of every model id, such as glm/glm-5.3, so it cannot change after the endpoint is created.

Preset endpoints

Preset	Base URL	Notes
GLM (Z.AI)	https://api.z.ai/api/anthropic
Qwen (Alibaba Model Studio)	https://coding-intl.dashscope.aliyuncs.com/apps/anthropic	Coding Plan. Pay-as-you-go uses https://{WorkspaceId}.<region>.maas.aliyuncs.com/apps/anthropic
Kimi (Moonshot)	https://api.moonshot.ai/anthropic
DeepSeek	https://api.deepseek.com/anthropic
MiniMax	https://api.minimax.io/anthropic	Mainland China: https://api.minimaxi.com/anthropic
Ollama (local)	http://localhost:11434	Any API key works. No models are pre-filled; context window starts at 65536. See below

Vendors rename models often. If a turn fails with an unknown-model error, check the vendor’s documentation and edit the model list.

Local models with Ollama

Ollama serves the Anthropic Messages API on your own machine, so a local model can drive the same agent.

1. Pick the Ollama (local) preset. The base URL is http://localhost:11434 and the API key is pre-filled with ollama. Ollama ignores the key, so any value works.
2. Add each model by the exact name ollama list prints, for example qwen3.6:35b-a3b-coding or hf.co/user/repo:Q4_K_M.
3. Set Context window to the OLLAMA_CONTEXT_LENGTH your Ollama server runs with. The preset starts at 65536. A coding agent needs at least 64k tokens to hold its tool definitions and a working conversation.

Kanna maps the agent’s Claude model names (Haiku, Sonnet, Opus, and the model used for subagents and background tasks) to the model you chose, so those calls do not ask Ollama for a Claude model it does not have.

Context window

Leave Context window empty and the agent assumes 200,000 tokens for any non-Claude model. If the real window is smaller, the conversation outgrows it and the endpoint rejects or truncates the request. Setting the field makes the agent compact the conversation before it reaches that size, and the composer’s context meter shows the same limit. The value is per endpoint, from 4,096 to 2,000,000 tokens.

Chat with it

In the chat composer, open the provider menu and choose Custom endpoints. The model picker lists every model of every endpoint you added, labelled with the endpoint, for example GLM (Z.AI) · glm-5.3. The entry only appears once at least one endpoint is configured.

Switching to a model on a different endpoint restarts the agent session on the new endpoint; the conversation history is kept. Subagents can use custom endpoints too: pick Custom endpoints as their provider in Settings → Subagents.

Your API key

The key is stored only in Kanna’s settings file on the machine running Kanna, and is sent only to the endpoint you configured. The browser never receives it: Settings shows just its last four characters. To change it, edit the endpoint and type a new key; leaving the field blank keeps the current one.

Limits

• No cost tracking. Kanna has no prices for these models, so turns on a custom endpoint report tokens but no dollar cost.
• No reasoning-effort controls, and no per-chat context-window picker. The vendor decides how its model maps effort; set it in the vendor’s console if it offers it. The window is a per-endpoint setting (see above).
• Claude model names are mapped to your model. For a custom endpoint Kanna points the agent’s Haiku, Sonnet, Opus and subagent model names at the model you picked. Switching to another model mid-chat restarts the agent session so background tasks follow the new model too; the conversation is kept. On Ollama this also means only one model needs to be loaded at a time.
• A stalled endpoint is failed after a timeout. If the endpoint accepts the request but never streams a reply, Kanna stops the turn and names the endpoint in the error, the same protection OpenRouter turns get.
• Deleting an endpoint that an existing chat uses moves that chat to the first remaining custom model on its next message. If none remain, sending fails with an error asking you to pick another model.