dsh-zhipu

Getting Zhipu BigModel web tools into DeepSeek Harness used to mean wiring up MCP endpoints one by one, hand-mapping tool names across providers, and juggling several API keys. dsh-zhibu collapses all of that into a single install: its bundled cordis.patch.yml selects the web_search and web_fetch providers, injects the ZAI_API_KEY environment variable, and sets a 60-second read budget. A web-service restart is the last step you take. One GLM Coding Plan key covers search, page reading, vision, and image transcription end to end.

Each capability is purpose-built. Web search runs on the search_pro engine, returns a fixed ten results with a maxResults truncation flag. Web reading renders pages server-side so JavaScript-heavy sites yield readable text, and the local machine never hits the target URL directly. Vision calls GLM-4.6V across eight task modes including OCR transcription, chart interpretation, and video-frame analysis, mirroring the official vision MCP system prompts. Image transcription is an agent pre-step hook: images dropped into a conversation are converted into fenced text descriptions before the message is queued, letting plain-text models consume them transparently; if the description call fails the original image block is preserved rather than silently dropped.

If you already subscribe to GLM Coding Plan, run text-only or lightweight multimodal models in DSH, and want search, page parsing, and image understanding on a single auth and billing rail, dsh-zhipu is the bridge that plugs the full Zhipu suite into your Harness.