---
title: "DeepSeek Harness coding plan & API key"
description: "DeepSeek Harness (dsh) on a coding plan from $10/mo: one provider in settings.yaml for Sonnet 5.5, Opus 5.5 and DeepSeek V4.1 Flash. No surprise charges."
url: https://codingpass.ai/tools/deepseek-harness
language: en
updated: 2026-11-17
---
# The coding plan for DeepSeek Harness.

Keep dsh exactly as it is. Add CodingPass as a provider with baseURL in settings.yaml and run Claude Sonnet 5.5, Opus 5.5 and DeepSeek V4.1 Flash on a monthly coding plan from one allowance.

- Setup in 60s
- No surprise charges
- Full refund if you've used 5% or less

~/.dsh/settings.yaml

```
llm-pi-ai:
  providers:
    codingpass:
      apiKeyEnv: CODINGPASS_API_KEY
      api: openai-completions
      baseURL: https://codingpass.ai/v1
      models:
        - id: claude-sonnet-5-5
        - id: deepseek-v4.1-flash
```

Model ID for your plan

**Go**

claude-sonnet-5-5

**Plus**

claude-opus-5-5

**Pro**

claude-fable-5-2

## A plugin-first agent harness.

DeepSeek's open-source agent harness, MIT licensed. All of it runs on CodingPass.

### Everything is a plugin

Tools, model routes and even the interface are plugins on the Cordis framework. Swap one out, install one from the community, or write your own.

### Claude and DeepSeek in one agent

Every provider's models become routes you switch under Settings → Models, no restart needed. Subagents can run on a different model than the main agent.

### Every step on record

Each session logs the requests, tool calls and subagents behind it, so you can trace what the agent did and on which model.

### A web UI, local or over SSH

`npx @deepseek-ai/dsh web` opens a web UI in your browser. Started over SSH, it serves the remote machine. Desktop builds cover macOS and Windows.

## Which plan for DeepSeek Harness?

Every plan works with dsh. Pick the Claude flagship you want as your default; all open-weight models, DeepSeek included, come with every plan and share the same monthly allowance.

- [Full plan details](https://codingpass.ai/pricing#compare-plans)
- [Allowances for all 24 models](https://codingpass.ai/models)

### Go: Claude Sonnet 5.5

Best for Daily coding, with DeepSeek for the routine work

$10/mo

$60 Sonnet allowance a month

**Requests / month**

≈2,050

**Per 5-hour window**

≈342req

**Each open-weight model**

$40/mo

**Model ID**

claude-sonnet-5-5

### Plus: Claude Opus 5.5

Best for Large refactors, long agent sessions

$30/mo

$200 Opus allowance a month

**Requests / month**

≈3,420

**Per 5-hour window**

≈570req

**Each open-weight model**

$125/mo

**Model ID**

claude-opus-5-5

### Pro: Claude Fable 5.2

Best for Parallel agents and the hardest problems

$100/mo

$500 Fable allowance a month

**Requests / month**

≈5,090

**Per 5-hour window**

≈848req

**Each open-weight model**

$300/mo

**Model ID**

claude-fable-5-2

Estimates from CodingPass traffic: about 100K tokens per request at roughly 5% input, 1% output and 94% cache read. A single dsh task usually makes several requests. Allowances are metered at official API prices.

## Three steps. About a minute.

dsh reads custom providers from `~/.dsh/settings.yaml`. Add CodingPass once and its models show up under Settings → Models.

1. ### Create a key
   Sign in with Google, GitHub or email and pick a plan. Then go to Dashboard → Keys → New key. Keys start with `cp-live-`.
2. ### Add the provider
   Paste the block into `~/.dsh/settings.yaml`, then export your key in `~/.zshrc` or `~/.bashrc` and open a new shell. Or add the same values as a custom provider under Settings → Models.
   `export CODINGPASS_API_KEY="cp-live-…"`
3. ### Start the web UI
   Open the address it prints, add your project as a workspace and pick a CodingPass model. On Pro, add `claude-fable-5-2` to the models list. [Get your request ID](https://codingpass.ai/tools#check).
   `npx @deepseek-ai/dsh web`

~/acme-crm Example session dsh web

```
$ npx @deepseek-ai/dsh web
  Web UI http://127.0.0.1:3080

> Merge duplicate contacts on import and keep the newest email.

● Plan 3 steps
● Edited src/import/contacts.ts (+38 −9)
● Ran pnpm test · 27 passing
  Duplicates match on normalized email; the newest record wins.
```

Sonnet 5.5

$2.80 / $60.00 resets in 11d

## Get more out of every allowance.

One monthly allowance covers every model, with 5-hour and weekly caps. The right model for the task makes it go further.

### What each model can use on Go

On Go, Sonnet 5.5 can use up to $60 a month, and each of the 8 open-weight models up to $40.

Go · up to, per model/mo ≈ req

Claude Sonnet 5.5 $60 ≈2,050

DeepSeek V4.1 Flash 1017 $40 ≈13,600

Kimi K2.7 Code $40 ≈1,570

GLM-5.3-Flash $40 ≈10,100

+ 4 more open-weight models, up to $40 each.

### Claude plans, DeepSeek writes the code

Keep Sonnet 5.5 (Opus 5.5 on Plus) as the main agent to plan the change and review the result, and let subagents on DeepSeek V4.1 Flash write the code. Both draw on the same allowance, and DeepSeek costs less per request.

Under Plugins → Subagent, turn on Model selection and add `deepseek-v4.1-flash`. New sessions can then put subagents on it; ask for the split in your prompt.

```
Plan this change first. Hand each step to a subagent on
deepseek-v4.1-flash, then review its diff before you finish.
```

## Questions about DeepSeek Harness.

**Is this allowed? Could my account get banned?**

You use only your CodingPass key. You never log in to, share or relay anyone's account, so your own Claude or ChatGPT account isn't involved. Your requests follow each vendor's usage policy.

**Why the OpenAI-compatible endpoint?**

dsh talks to custom providers over Chat Completions. `https://codingpass.ai/v1` serves every model on your plan that way, Claude included, so one provider block covers them all.

**Does it work with the desktop builds?**

Yes, if the build lets you add a custom OpenAI-compatible provider. Use the same three values: `https://codingpass.ai/v1`, your key and a model ID from your plan.

**What happens when I hit the 5-hour cap?**

Each 5-hour window allows 1/6 of the monthly allowance (on Go, $10 of Sonnet 5.5, about 342 requests). The cap counts all models together, so CodingPass then returns 429 with the reset time for every model. Wait for the reset, buy a reset card for the current window, or upgrade. [How caps and 429s work](https://codingpass.ai/usage-limits#response).

## Same key, every agent.
