modelspoke

modelspoke

English | 中文

A plugin for DeepSeek Harness for managing connections to local, OpenAI-compatible model servers. It is an improved replacement for the stock dsh custom provider setup: it discovers your servers and models instead of requiring hand-written provider blocks and per-model fields, and provides the capabilities the stock setup lacks.

Features

First-class llama-swap and Ollama support

The router llama-swap is highly recommended for local model hoarders since it can serve as a source of truth for model capabilities across harnesses. modelspoke understands the capability data that llama-swap adds to its extended OpenAI-compatible endpoint. Capability discovery also supports Ollama's API extensions.

Presets for common models

To help out with endpoints that don't have full capability discovery (e.g. llama-server, vLLM, sglang) modelspoke includes a table of capabilities for common base models to use in initial configuration.

Full-featured setup UI

Setup is easy to use and covers all the day-to-day fields (the deep template-contract fields — compat — stay hand-edited in the file). Allows for overriding presets and discovered capabilities and maintaining multiple setting profiles of the same underlying model.

Reasoning effort levels

dsh custom provider functionality doesn't afford any way to set reasoning effort. Modelspoke can discover the supported effort levels and allows customizing the map from the UI effort setting to the model supported setting.

Image input

Models with multimodal capabilities are great but if you add them via the dsh custom provider setup that functionality is not available. Modelspoke can discover image input capability or allow you to specify it. dsh 0.1.5 renders the read_image tool's inline results in session chat natively — modelspoke's role is surfacing which models accept images at all.

Installation & setup

Prerequisite: dsh 0.1.7 (verified against 0.1.7-rc.1) with the dsh plugin command.

1. Install: dsh installs plugins with a tool called pnpm. If you don't have it yet, install it first:

   npm install --global pnpm
   

   Then one command for each profile where you will use local models:

   dsh plugin --profile web add modelspoke
   dsh plugin --profile headless add modelspoke
   

2. Restart dsh if it is running, so it picks up the plugin.

3. Open the Modelspoke settings card. In the dsh web UI, select Plugins in the left sidebar, open the modelspoke row in the Installed group — the Modelspoke card renders on its bundle detail page, then + Add provider:

   

4. Point it at your server. Set the provider's name, its base URL, the environment variable name holding the API key (omit for keyless local servers — no auth header is sent in that case), and an optional default effort (minimal … max) — then commit with the card's Apply button. The row's status dot goes green once the model fetch succeeds.

5. Configure per model where you want to. Expanding a provider fetches its model list; each model row has a chevron that opens an editable detail (context window, max output tokens, the thinking-level map, nothink, image input, reasoning effort):

   

   The model list is the curation — a model is addressable by the agent only while it is in the list; clearing a detail field releases that field back down the resolution chain.

Appendix

• docs/usage.md — using modelspoke after install: the resolution chain, the per-model detail, nothink models, images, and the profile patch configuration shape
• docs/preset-authoring.md — authoring a model preset from the template in the artifact (the preset-draft / drift-check workflow)
• docs/llama-swap-setup.md — the minimal llama-swap setup, and how modelspoke reads llama-swap's extended endpoint
• docs/design.md — architecture and decisions
• docs/provider-details.md — the provider reference: why the five backends, where each capability value comes from, per-provider quirks
• docs/dsh-plugin-guidance.md — integrating with dsh: the adapter registration contract, the web-UI half, settings writes, and the read_image tool-view history