Updated November 2, 2026

DeepSeek SuperAgent: GitHub, Setup, Model Providers & Costs

By Cameron Voss and West Mao

A practical guide to the research prototype: what it does, how to get started, what model calls may cost, and how it compares with GitHub Copilot's cloud agent.

What is DeepSeek SuperAgent?

DeepSeek SuperAgent is a research prototype for handling coding tasks in a GitHub repository. An authorized teammate can address the agent in GitHub, and the prototype creates a work item, starts an execution session and replies in the same conversation. It can also receive requests through Feishu.

There are two related but distinct things: the Super-agent design article, which discusses a broader architecture, and the public dsh-superagent-prototype, which implements a subset of those ideas. The public code differs from both the team's internal implementation and the design article. Some ideas in the article have not been implemented.

How does the SuperAgent architecture work?

A request can arrive through GitHub or another configured source. The system checks the request against its policy and decides whether to create work. A main session coordinates the work; executor sessions handle assigned tickets in separate directories and report back to the original conversation.

A person still needs to review the work. The setup guide recommends leaving automatic merging off while the team learns how the agent behaves.

Where is the official SuperAgent GitHub repository?

The official source is deepseek-ai/dsh-superagent-prototype. The code is released under the MIT license; keep the required copyright and license notices when using or redistributing it.

For installation, use the repository's SETUP.md and deployment guide. They may change as the prototype and Harness evolve.

How do you set up DeepSeek SuperAgent?

Before you start, prepare a machine you control, a test repository, a separate GitHub account for the agent and a model provider. The setup guide lists Node.js 22.19+ or 24+, pnpm, Git 2.32+, tmux, Python 3.8+, GitHub CLI 2.48+ and a compatible DeepSeek Harness version. Feishu requires additional components.

1. Prepare an isolated environment. Start with a test repository, a dedicated agent account and limited repository permissions. Choose who is allowed to request work.
2. Install Harness and the prototype. Follow the exact version and commands in the current deployment guide; install and build the bundle, then add it to a dedicated SuperAgent profile.
3. Configure the profile and model provider. Set the required machine configuration and connect a working model route in that profile. Keep credentials out of issue comments and shared logs.
4. Connect GitHub and set policy. Configure the repository, agent identity and permitted users. Start with GitHub only if you do not need Feishu.
5. Run one small test. Ask the agent to make a bounded change in the test repository. Check the task record, resulting diff, logs, tests and model usage before expanding its access.

Upgrade warning: the official profile-creation command replaces an existing profile and can delete its configuration. It is for a new installation, not a routine upgrade. Consult the deployment guide before re-running it.

How do you configure a model provider for SuperAgent?

SuperAgent uses its Harness profile's default model unless the evaluation or execution settings select another. Configure the provider in the profile that runs the agent and check the deployment guide for the current profile and panel settings.

Option 1: Use the official DeepSeek API

In the Harness Web UI, open Settings → Models, enter your DeepSeek API key on the DeepSeek card and save it. Select an available model and run a small test. Provider setup and SuperAgent setup are separate steps; make sure the credentials and model are available to the profile actually running the agent.

Option 2: Use another provider or a custom gateway

Harness documents built-in provider options and a Custom model API option. For a custom route, you supply its provider ID, base URL, supported protocol, credential and model ID. Match the protocol to the endpoint and test a simple model call before assigning an agent task. A model appearing in a configuration list does not guarantee the endpoint accepts every request.

How much does DeepSeek SuperAgent cost to run?

The code is open source, but model calls and the machine running the agent can cost money. Your bill depends on the model, how much context each request includes, cache usage, output and the number of worker sessions. A task that needs several retries can cost more than a straightforward one. Measure your own workload before estimating a monthly bill.

Official DeepSeek API price example

The following is an illustrative official API calculation, not a Coding Plan price and not a measured SuperAgent task. As checked November 2, 2026, the official pricing page lists deepseek-flash cache-miss input at $0.30 per million tokens and output at $1.20 per million tokens during peak hours. Cache-hit input is $0.006 per million tokens. Off-peak prices are lower; check the provider's current schedule and rates.

Calculate cache-hit input separately at the applicable rate. The example excludes hosting and other charges, and a real task may use more or fewer tokens. Coding Plans have their own usage windows and limits; this API calculation does not describe those plans.

How can you control usage?

• Start with one small, well-scoped task and record the provider's usage before and after.
• Limit concurrent workers and check logs for retries or repeated long-context calls.
• Use the provider's budget or alert controls where available; check the balance and limits of your chosen plan.
• Compare the cost of a useful, reviewed result with the work it saved, rather than looking only at the cost of one request.

DeepSeek SuperAgent vs. GitHub Copilot cloud agent: what is different?

This is a workflow comparison, not a benchmark of coding quality. GitHub's features, model availability and allowances can change.

Frequently asked questions

Is DeepSeek SuperAgent free?

The prototype's source code is available under the MIT license. Running it may still cost money for models, a machine or server, and other services.

Is SuperAgent a DeepSeek model?

No. It is a coding-agent prototype built on DeepSeek Harness. It needs a separately configured model provider to perform model calls.

Does it require a DeepSeek API key?

A configured model provider is required; the Harness documentation also describes third-party and custom providers. Whether a particular provider works for your workflow depends on its protocol, model and the version you use.

Can I use it with an existing Coding Plan?

Possibly, but compatibility between this prototype and our Coding Plans has not been verified here. Check the current endpoint, model route, plan rules and actual behavior before relying on it.

Should I let it merge pull requests automatically?

Review and merge changes yourself at first. The official setup recommends keeping auto-merge off while you learn how the agent behaves.

Prices and documentation can change. DeepSeek V4 Hub is an independent site and is not affiliated with or endorsed by DeepSeek or GitHub.