Chinese-model coding CLIs
September 15, 2026 · View on GitHub
Taskuary includes Qwen Code, OpenCode (DeepSeek, GLM, MiniMax), and Kimi Code (Moonshot AI) under Connections → AI CLI agents. Update and restart Taskuary if those entries are missing.
These are coding tools running on your machine. A cloud model still receives the context sent to it. Provider accounts, pricing, and regional availability are separate from Taskuary; this integration does not establish mainland-China availability.
DeepSeek through OpenCode
DeepSeek's official integration guide recommends OpenCode as one way to use its models. Taskuary runs OpenCode, not an official standalone "DeepSeek CLI."
- Open Connections → AI CLI agents → OpenCode (DeepSeek, GLM, MiniMax) and click Install.
Windows uses
npm install -g opencode-ai, so Node/npm must be installed. macOS and Linux also have a standalone installer. - Click Set it up. In OpenCode, type
/connect, choose DeepSeek, and enter your DeepSeek API key directly in the CLI. - Use
/modelsto select an available DeepSeek model. - Save the connection, click Test, then choose its coding worker for a task.
Leave Taskuary's model field blank to use OpenCode's selection, or enter an
exact
provider/modelidentifier fromopencode models.
The same OpenCode connection can use Z.AI's GLM or MiniMax: select the
appropriate provider with /connect, then choose its model with /models.
Regional and coding-plan providers may use different credentials or endpoints;
follow OpenCode's provider instructions.
Kimi Code
- Open Connections → AI CLI agents → Kimi Code (Moonshot AI) and click Install.
Taskuary prefers Moonshot's standalone installer. The npm alternative is
npm install -g @moonshot-ai/kimi-codeand requires Node 22.19 or newer. Windows also requires Git Bash. - Click Set it up, run
/login, and complete Kimi or Moonshot authentication. - Save and Test, then assign the Kimi coding worker to a task. Leave the model blank to use Kimi's configured default, or enter a configured model alias.
This preset targets the current MoonshotAI/kimi-code
project. Upgrade old Python kimi-cli installations before using it; the command-line
flags differ. See the current command reference.
Qwen Code
Use the Qwen Code setup guide for installation, provider selection, local-model configuration, and native ACP support.
Roles and sessions
Taskuary's model picker includes provider-specific suggestions. Qwen reads model IDs
for the active protocol from QWEN_HOME/settings.json (normally ~/.qwen/settings.json);
its OAuth choice is coder-model. Kimi reads aliases from
KIMI_CODE_HOME/config.toml (normally ~/.kimi-code/config.toml), with its documented
/login aliases as the fallback. Those files are re-read when the picker loads.
OpenCode suggestions use exact provider/model IDs from its provider catalog;
connect the corresponding provider before selecting one. A blank model still keeps
the CLI's own default. Model choices never change your credentials or sign-in method.
Identifier references: Qwen model providers, Kimi model aliases, and OpenCode models.
- OpenCode and Kimi: coding tasks and general-agent tasks with tools enabled. Taskuary refuses to use these presets for message triage or read-only reports, because a per-run restriction on their inherited tools has not been verified. Use Qwen Code, Ollama, or an API provider for those roles.
- Headless runs stream readable text and tool activity, save the CLI session ID, and resume the same conversation. Interactive coding tasks open the real CLI. Automatic discovery of a new interactive pane's session ID is not implemented for these two CLIs; use the CLI's own session picker to reopen such conversations.
- Both vendors offer ACP, but these two Taskuary presets currently use their native command-line interfaces. Their ACP model selection and session behavior are not yet part of Taskuary's verified integration.
- Kimi takes headless prompts as an argument, so the operating system's command-line length limit applies. OpenCode reads headless prompts from stdin.
Compatibility evidence
Verified on Windows on 2026-09-15 using OpenCode 1.18.31 and Kimi Code 0.43.1, with isolated CLI homes and a local mock model endpoint:
- Prompt delivery and readable output through Taskuary's actual runner.
- Session IDs and resumed requests containing the previous conversation.
- A real file write requested by the mock model, plus structured tool progress.
- Focused regression tests for connection adoption, installers, updates, interactive arguments, model forwarding, output parsing, and restricted-role rejection.
No paid DeepSeek, GLM, MiniMax, or Kimi service was called. Real-provider authentication, billing, model quality, and mainland-China connectivity still need user-side verification.