# DeepSeek Harness

Canonical page: https://agentsky.dev/agents/dsh

DeepSeek's official first-party coding agent in the cloud.

DeepSeek Harness (DSH) is DeepSeek's first-party agentic coding assistant, tuned to work natively with its verified DeepSeek, Kimi, GLM, and Grok model set. It is the only harness purpose-built for DeepSeek's model architecture and correctly handles DSH's SSE tool-name assembly, which generic wrappers do not.

## Facts

- **Agent type:** `dsh`
- **Default model:** DeepSeek V4 Pro (`deepseek-v4-pro`)
- **Start in the browser:** https://agentsky.dev/playground/enter?agentType=dsh&llm=deepseek-v4-pro&source=agent-md
- **API reference:** https://agentsky.dev/docs
- **Billing:** Running sessions are billed by compute time and model tokens; parked agents incur no compute charge.

## Models it runs on

USD per 1M tokens.

| Model | Id | Input | Output |
|---|---|---|---|
| DeepSeek V4 Pro (default) | `deepseek-v4-pro` | $0.924 | $2.772 |
| DeepSeek V4.1 Flash | `deepseek-flash` | $0.21 | $0.84 |
| Kimi K3.1 Pro | `kimi-k3.1-pro` | $2.20 | $11.00 |
| Kimi K3 | `kimi-k3` | $3.30 | $16.50 |
| Kimi K2.7 Code | `kimi-k2.7-code` | $1.045 | $4.40 |
| GLM-5.3 | `glm-5.3` | $1.54 | $4.84 |
| GLM-5.3-Flash | `glm-5.3-flash` | $0.165 | $0.55 |
| Grok 4.6 | `grok-4.6` | $2.20 | $6.60 |
| GPT-6 Astra | `gpt-6-astra` | $7.00 | $35.00 |
| GPT-6.1 Sol | `gpt-6.1-sol` | $1.40 | $7.00 |
| GPT-6 Sol | `gpt-6-sol` | $1.40 | $7.00 |
| GPT-6 Luna | `gpt-6-luna` | $0.07 | $0.35 |
| GPT-5.6 Sol | `gpt-5.6-sol` | $3.50 | $21.00 |
| GPT-5.6 Terra | `gpt-5.6-terra` | $1.40 | $8.40 |
| Claude Opus 5.5 | `claude-opus-5-5` | $2.80 | $14.00 |
| Claude Opus 5 | `claude-opus-5` | $3.50 | $17.50 |
| Claude Fable 5.1 | `claude-fable-5-1` | $7.00 | $35.00 |
| Claude Sonnet 5 | `claude-sonnet-5` | $1.40 | $7.00 |
| GPT-5.6 Luna | `gpt-5.6-luna` | $0.14 | $0.84 |
| Claude Sonnet 5.5 | `claude-sonnet-5-5` | $1.40 | $7.00 |
| Claude Haiku 5.5 | `claude-haiku-5-5` | $0.07 | $0.35 |
| Claude Haiku 4.5 | `claude-haiku-4-5` | $0.70 | $3.50 |
| Gemini 4 Argon | `gemini-4-argon` | $4.40 | $22.00 |
| Gemini 3.8 Flash | `gemini-3.8-flash` | $1.65 | $8.25 |
| Gemini 3.5 Flash-Lite | `gemini-3.5-flash-lite` | $0.33 | $2.75 |
| Grok 4.7 | `grok-4.7` | $2.20 | $6.60 |
| Muse Spark 1.3 | `muse-spark-1.3` | $1.375 | $4.675 |
| Muse Spark 1.2 | `muse-spark-1.2` | $1.375 | $4.675 |
| Muse Spark 1.1 | `muse-spark-1.1` | $1.375 | $4.675 |
| Muse Wonder 1.0 | `muse-wonder-1.0` | $4.40 | $22.00 |
| MiniMax-M3 | `minimax-m3` | $0.33 | $1.32 |
| MiniMax-M3.1-Flash | `minimax-m3.1-flash` | $0.165 | $0.66 |

## What it can do

- **DeepSeek-native:** Optimised for DeepSeek V4 Pro and Flash — lowest cost per task in the catalog.
- **Research & analysis:** Deep reasoning on long documents, code, and data.
- **Code generation:** Writes and refactors code with chain-of-thought reasoning visible.
- **Multi-model:** Also runs Kimi K3, GLM-5.3, and other compatible models.

## Questions

### How do I run DeepSeek Harness in the cloud on AgentSky without setting up my own server or Docker container?

Create an AgentSky account and start a DSH agent via the API or web UI. AgentSky provisions a persistent cloud computer, installs DeepSeek Harness, and runs your task. Specify the model and task in the same call. No Docker, no server setup, no dependency maintenance.

### Which models work with DeepSeek Harness on AgentSky, and why are GPT and Gemini models excluded from the model list?

DSH on AgentSky supports DeepSeek V4 Pro, DeepSeek V4.1 Flash, Kimi K3.1 Pro, Kimi K3, Kimi K2.7 Code, GLM-5.3, GLM-5.3-Flash, Grok 4.6, GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna, GPT-5.6 Sol, GPT-5.6 Terra, Claude Opus 5.5, Claude Opus 5, Claude Fable 5.1, Claude Sonnet 5, GPT-5.6 Luna, Claude Sonnet 5.5, Claude Haiku 5.5, Claude Haiku 4.5, Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.5 Flash-Lite, Grok 4.7, Muse Spark 1.3, Muse Spark 1.2, Muse Spark 1.1, Muse Wonder 1.0, MiniMax-M3, and MiniMax-M3.1-Flash. GPT models break on DSH's SSE tool-name assembly (verified live 2026-08-14), and Gemini's compatibility plane returns 400 errors on DSH's request shape. This is the verified working set and the complete current selection.

### What does it cost to run DeepSeek Harness on AgentSky per task, and how is it billed?

DeepSeek V4.1 Flash is the fastest and lowest-cost model option for DSH; DeepSeek V4 Pro offers stronger reasoning at higher per-token cost. Playground coverage for DSH is expanding, so task-level median costs are not yet published. Current per-model rates are on the pricing page.

### My DeepSeek Harness agent keeps stopping in the middle of a coding task before it finishes — what is the usual cause?

The most common stop causes for DSH are: a shell command exit code that DSH treats as terminal, a context window overflow on a large codebase, or a model response that returns a tool shape DSH does not expect. Check the task log for the last tool call. Using DeepSeek V4 Pro over V4.1 Flash reduces mid-task stops on complex tasks due to its larger context and stronger reasoning.

### If my DeepSeek Harness agent is mid-task and the AgentSky machine it runs on is restarted or replaced, does it lose its coding progress?

The session workspace persists across ordinary restarts. If the underlying machine is replaced, AgentSky restores the latest durable recovery point, preserving the working directory and recent file edits. Work in progress at the exact moment of a hard replacement may replay from the last checkpoint. Committing intermediate changes with git provides a durable recovery layer beyond AgentSky's managed checkpoints.

### How does running DeepSeek Harness on AgentSky compare to cloning the repo and running it on my own development machine?

Self-hosting DSH means managing the runtime environment, keeping dependencies in sync with upstream DeepSeek releases, and handling crash recovery and uptime yourself. On AgentSky, the harness is maintained, the runtime is managed, and the agent gets a persistent cloud computer that survives crashes and laptop closure.