Set up DSH (DeepSeek Harness)

01 / PREPARE

Install the agent and prepare access

1. Install the agent using its official guide. Check its version before merging the examples below.
2. Purchase your Multi-Model plan, then open Billing & API keys to copy your active key. OpenAI base: https://multi-api.savemytoken.com/v1. Claude Code base: https://multi-api.savemytoken.com. DeepSeek credentials are a separate scope.
3. Back up the file you will edit. Merge fields into existing objects and tables; preserve unrelated providers, tools and permission settings. Keep credentials in private local settings.
4. Image input is supported by seven routes; GLM-5.3 is text-only. Send JPEG, PNG, GIF or WebP images as Base64 or public HTTP(S) image URLs. The gateway accepts up to 20 images, 10 MiB per inline image and 20 MiB total inline image data per request.
5. Each image temporarily reserves 16,384 input tokens at the selected model's points rate. After completion, the reservation is replaced by the supplier's actual input and output token usage, including image tokens. Base64 characters are not billed as text tokens.
6. Start dsh web, open the printed local address, then Settings → Models → Add model provider → Custom model API. Select OpenAI Chat Completions and enter the delivered /v1 base and key in that form.

Ready when: the agent opens, and you know which plan, process and configuration directory you will use.

02 / CONFIGURE

Save the connection and model catalog

Create the parent directories if needed. Download or copy each file, replace the example endpoint and local placeholders, then merge it into the stated location. Downloads and the examples below contain the same configuration.

- id: llm-pi-ai
  config:
    providers:
      savemytoken-multi-model:
        compat:
          supportsDeveloperRole: false
        models:
          - id: deepseek-flash
            reasoningEfforts:
              off: none
              low: low
              high: high
              max: max
          - id: glm-5.3-flash
            reasoningEfforts:
              low: low
              high: high
              max: max
          - id: mimo-flash
            reasoningEfforts:
              off: none
              high: high
          - id: mimo-pro
            reasoningEfforts:
              off: none
              high: high
          - id: glm-5.3
            reasoningEfforts:
              low: low
              high: high
              max: max
          - id: kimi-k3
            reasoningEfforts:
              low: low
              high: high
              max: max
          - id: qwen3.8-max
            reasoningEfforts:
              off: none
              low: low
              medium: medium
              xhigh: xhigh
          - id: minimax-m3
            reasoningEfforts:
              off: none
              high: high

Ready when: the files parse successfully and the selected profile points to the delivered URL, key reference and exact model catalog.

03 / START

Start the agent and choose your model

Open a terminal in a trusted project. Complete the file setup before running these commands.

dsh web

Use Provider ID savemytoken-multi-model in the form. Fetch available models, select the permitted IDs below and save. If discovery is unavailable, add each exact ID and display name manually. Choose the provider beside the composer.

Check: the provider is the one you configured and the selected model is deepseek-flash. A visible model list alone does not prove inference works.

04 / REASONING

Choose and verify reasoning effort

Model discovery can supply names without effort metadata. Merge the capability fragment into the active profile, restart and open the Effort menu. Off explicitly maps to none; a null/empty mapping may only omit the parameter.

This patch does not set the starting effort. Before the first request, select the default listed in the table: High for DeepSeek, Max for GLM/Kimi, XHigh for Qwen, and High (On) for MiMo/MiniMax. Recheck after switching models.

Choose provider → choose model → choose Effort → start a new session

On DeepSeek Flash, choose Low, then High, then Low again. Inspect the active setting each time before sending the next message. A longer answer is not evidence that effort changed.

REASONING CONTROLS

Choose effort for the actual model

Use these exact gateway IDs. All eight models draw from the same points allowance. Image input support is listed per model below. Output is text or tool calls; audio and video input and media generation are not included.

05 / DESKTOP

Harness Desktop / local Web UI

Harness Desktop has the same custom-provider setup flow, but can use a different profile from dsh web. Add the provider and credentials in Desktop's own Settings → Models.

Use the app's configuration-file action to locate the active profile before merging model capabilities. Editing the web profile does not configure Desktop. Restart and choose the intended model and effort in a new session.