dsh-iris

Tools, Skills, and MCP servers pile up inside DeepSeek Harness, and a model staring at a wall of capability cards tends to lose its way. The usual fixes add a routing model for secondary dispatch or preload every Provider at the metadata stage, both of which inflate permissions and cold-start cost. dsh-iris turns capability exposure into a progressive gate: the current model reads short descriptions, picks one capability, and then follows native loading and execution in DSH. No extra routing model, no GPT-6 API dependency.

Small catalogs surface capabilities as budgeted cards directly; larger ones keep an iris_search long-tail search available. Native tools, already-mounted Skills, and MCP servers continue on their own routes without redundant activation through Iris. Metadata reads never import Providers, load Skill bodies, or connect MCP servers, and description text carries no permissions. Four modes-Minimal, Standard, Code, and Cordis-each respect their own boundary while preserving the native persona and the host approval chain.

Built for developers who have wired multiple tool families or MCP servers into DeepSeek Harness and want the model to select the right capability on demand, without adding a routing layer or a second Skill/MCP runtime. Iris does not install community plugins and does not proxy business calls; it does one thing: put the right capability in front of the model at the right moment.