Preface¶

The philosophy of DSH is “everything is a plugin,” and agents can acquire external capabilities through plugins. For scenarios that require web search, image understanding, audio/video processing, and speech synthesis/recognition, the common approach is to manually assemble API requests in the shell, handle base64 payloads, and parse responses; large payloads can also be affected by shell capture limits. dsh-mimo-agent-tools encapsulates Xiaomi MiMo’s search and multimodal capabilities as DSH agent tools, reducing such manual workflows.

What This Is¶

dsh-mimo-agent-tools is a DSH plugin that provides Xiaomi MiMo search + multimodal tools for DSH agents. The core tools include mimo_search, mimo_vision, mimo_audio, mimo_video, mimo_asr, and mimo_tts.

The plugin name in package.json is dsh-mimo-agent-tools, version 0.1.0, licensed under MIT; the README title uses mimo-agent-tools. The maintainer in the plugin listing is ch1bug.

Core Features¶

Search and Reasoning¶

• mimo_search: Performs web search via the Xiaomi MiMo API, returning answer + cited sources (including site_name/logo_url); optional force_search, user_location.
• mimo_think: Deep reasoning, returning the full reasoning chain reasoning_content + final answer.
• mimo_json: Outputs structured JSON via response_format: json_object.

Multimodal Understanding¶

• mimo_vision: Image understanding, supports local files or public URLs, and multi-image input.
• mimo_audio: Audio understanding/transcription, supports wav/mp3/flac/ogg/m4a.
• mimo_video: Video understanding, supports mp4/webm/mov, with optional fps / media_resolution.

Speech Synthesis and Recognition¶

• mimo_asr: Speech-to-text, with optional language hint.
• mimo_tts: Text-to-speech, outputs .wav/.mp3, supports preset voices or voice design, with optional style / format.
• mimo_voiceclone: Voice cloning, using reference clip + text, with optional format.

Usage Guidance¶

• audio-tools skill: Guides when to use mimo_asr, mimo_tts, mimo_voiceclone, mimo_audio. These tools are always registered; the skill only teaches usage and does not gate the tools.

Installation and Enablement¶

First, confirm the runtime environment:

1. The DSH host provides shell and sandboxPolicy services.
2. python3 is available on the host.
3. driver/mimo_driver.py exists.
4. A MiMo API key is ready; web_search also requires the connected-search plugin to be enabled in the MiMo console.

The API key is resolved at tool-call time:

1. Priority is given to reading XIAOMI_API_KEY from the DSH credentials service.
2. Falls back to the environment variables XIAOMI_API_KEY or MIMO_API_KEY.

The plugin does not hardcode keys.

Install the DSH bundle from a local directory:

dsh plugin --profile web add /path/to/dsh-mimo-agent-tools

Install the Python driver to the default path:

mkdir -p ~/.local/lib/mimo-agent-tools/driver
cp driver/mimo_driver.py ~/.local/lib/mimo-agent-tools/driver/

If the driver is not in the default path, you can configure it:

MIMO_DRIVER=~/.local/lib/mimo-agent-tools/driver/mimo_driver.py
MIMO_TMP=/tmp

Here, the default value of MIMO_DRIVER is ~/.local/lib/mimo-agent-tools/driver/mimo_driver.py, and the default value of MIMO_TMP is /tmp.

The plugin depends on @deepseek-ai/dsh-tools, declared as peerDependencies in package.json.

Typical Usage¶

Search¶

mimo_search can be used for web search. user_location can bias results, for example, by country/region/city dimensions. force_search is an optional parameter.

Images, Audio, Video¶

• mimo_vision: Pass local files (automatically base64-encoded) or public URLs; supports multiple images.
• mimo_audio: Used for wav/mp3/flac/ogg/m4a files.
• mimo_video: Used for mp4/webm/mov files; optional fps and media_resolution.

ASR / TTS / Voice Cloning¶

• mimo_asr: Optional language hint; do not pass the thinking field.
• mimo_tts: The target text goes in the assistant message; optional style / format.
• mimo_voiceclone: Provide a reference clip and text; optional format, which can be wav/mp3.

Reasoning and Structured Output¶

• mimo_think: In multi-turn tool conversations, the previous reasoning_content must be echoed back, otherwise the API may return a 400 error.
• mimo_json: The prompt needs a clear JSON shape description; max_completion_tokens must be sufficient to avoid JSON truncation mid-stream.

Applicable Scenarios and Notes¶

Suitable for letting agents in DSH call Xiaomi MiMo’s search, image, audio, video, ASR, TTS, and voice cloning capabilities.

Notes before use:

• web_search depends on the connected-search plugin enabled in the MiMo console.
• The plugin requires the DSH host’s shell and sandboxPolicy services, as well as host python3 and driver/mimo_driver.py.
• The Python driver is used to avoid shell stdout truncation; DSH’s bash seam caps at 64KB.
• mimo_think and mimo_json have specific prompt / multi-turn field requirements; use them according to the constraints above.
• The plugin runs with the current DSH process permissions; review the source code, dependencies, and MIT license before installation.

Conclusion¶

The value of dsh-mimo-agent-tools is bringing Xiaomi MiMo’s search and multimodal capabilities into the DSH agent toolchain: search, thinking, JSON, image, audio, video, ASR, TTS, and voice cloning are all exposed as tools, with a Python driver reducing the risk of handling large payloads in the shell.

Plugin directory listing:

https://www.skillhub.cn/plugins/ch1bug/dsh-mimo-agent-tools

GitHub repository listing:

https://github.com/ch1bug/dsh-mimo-agent-tools

The above links come from the plugin listing; the community directory is an independent site with no official affiliation to DeepSeek / High-Flyer, nor is it an official app store.