> ## Documentation index
> Fetch the complete index at: https://maritime.sh/llms.txt
> Setup procedure for AI agents (install, sign in, deploy, verify): https://maritime.sh/agents.md
> This page as HTML: https://maritime.sh/docs/frameworks/deepseek-harness

# DeepSeek Harness Guide

DeepSeek Harness (dsh) is DeepSeek's open-source agent harness, and Maritime runs it as a managed agent with one
command, on its own machine that sleeps when idle and wakes on the next message.

## Overview

[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) (dsh) is DeepSeek's open-source agent harness,
released in August 2026. Everything in it is a plugin: models, tools, skills, sessions, sandboxes, and the web
workbench itself. Every run is recorded in append-only session logs you can resume, fork, and replay from the
workbench.

Upstream is an early developer preview that promises breaking changes between releases, so Maritime pins the exact
dsh version in its template image and upgrades deliberately after testing each release.

## Docker Image

Image: `ghcr.io/maritime-sh/dsh-agent:0.1.0-rc.7`

A first-party Maritime image: Node 22, dsh pinned to the tag version, the web workbench on port `3080` behind a
per-agent auth layer, and agent state under `/data/.dsh` so sessions survive restarts and upgrades.

## How It Works

The upstream web workbench ships without authentication, so the Maritime image runs it loopback-only behind a small
auth shim that requires a per-agent token on every request. The Maritime dashboard injects that token for you: click
**Harness Workbench** on the agent page and you are in. The default model routes through the metered Maritime LLM
proxy, so a fresh agent works with no API key at all. Chat messages from Telegram, the CLI, and share links run
one-shot headless turns inside the agent's own container.

```
# The container runs with:
DSH_HEADLESS=true
DSH_PORT=3080
DSH_GATEWAY_TOKEN=<auto-generated>
DSH_MODEL=gpt-5-mini
DSH_HOME=/data/.dsh (baked in the image)
```

## Environment Variables

If you don't supply `OPENAI_API_KEY`, Maritime injects a per-agent proxy token pointed at
`https://api.maritime.sh/api/llm/v1`. Set `DSH_MODEL` to change the default model (any model the proxy serves).
You can also add your own providers, including DeepSeek's API, in the workbench under **Settings > Models**; keys you
add there are stored inside your agent's volume, never on Maritime's control plane.

## Deploy

One command with the [Maritime CLI](https://maritime.sh/docs/cli):

```
maritime create my-agent --template dsh
```

Or open **New agent** in the dashboard and pick the **DeepSeek Harness** template. New here? The
[quickstart](https://maritime.sh/docs/quickstart) covers accounts and the
[Templates guide](https://maritime.sh/docs/templates) covers cloning and sharing agents.

## Talk to it

Open **Harness Workbench** from the agent page for the full experience: sessions, plan mode, tool approvals, session
replay, and plugin management. Or chat from the terminal:

```
maritime chat my-agent "clone my repo and run the tests"
```

Terminal and channel chats are one-shot headless runs: each message boots the harness, runs the task to completion
with tools enabled, and returns the final answer. Recent conversation context carries over between messages;
long-lived interactive sessions live in the workbench.

### Upstream links & resources

- [deepseek-ai/deepseek-harness](https://github.com/deepseek-ai/deepseek-harness): the source repository (MIT
  license), including the Cordis plugin framework docs.
- [Official dsh documentation](https://deepseek-harness.github.io/deepseek-harness/en/guide/quickstart): quickstart,
  provider configuration, CLI profiles, and the plugin development guide.
- [dsh-plugin topic on GitHub](https://github.com/topics/dsh-plugin): community plugins you can install into your
  agent from the Console.
