Current low prices. Pro $29 → $35 Sep 17 ET.View prices
DeepSeek V4
DeepSeek V4
← All integrations
Terminal coding agentOpenAI Chat CompletionsExperimentalDocs: 0.1.1-rc.2Sources checked 2026-08-24

DeepSeek Harness Setup with DeepSeek V4

Configure the official DeepSeek Harness developer preview with DeepSeek-V4.io as a custom OpenAI-compatible provider.

Create API KeyRead API Docs

DeepSeek-V4.io is an independent API gateway, not the official DeepSeek website. A paid DeepSeek-V4.io key is required.

How the integration works

DeepSeek Harness is DeepSeek's official open-source, plugin-oriented agent runtime. Its local Web UI accepts a custom provider with a base URL, protocol, credential, and model catalog, so the text-only DeepSeek-V4.io gateway can be configured without exposing upstream provider credentials. DeepSeek has separately released an experimental vision model, but this gateway does not expose that model today. This site is independent and does not develop or represent the Harness.

Developer-preview boundary. DeepSeek labels Harness a developer preview and explicitly warns that compatibility-breaking changes will occur. The npm release moved from 0.1.0-rc.6 to 0.1.1-rc.2 in eight days, so pin the exact release candidate for repeatable tests and re-check this setup after every upgrade.
Vision is official, but not enabled through this gateway. DeepSeek now documents deepseek-v4-flash-vision-exp for image input, and the current Harness line can advertise image-capable models. DeepSeek-V4.io still exposes only deepseek-v4-flash, so keep this custom route text-only. Do not add image input or the vision model ID here until this gateway passes an end-to-end image, usage, and billing verification.
One site key

Create and manage it on /developer.

OpenAI Chat Completions

The client talks to its native compatible endpoint.

Upstream hidden

Provider credentials and routing stay on the server.

Before you start

  • Node.js installed
  • A paid DeepSeek-V4.io account with an API key from /developer
  • A local project directory you are willing to expose to an agent with file and shell tools

Safety before the first run

A coding agent operates with the workspace and execution permissions you grant it. Start with the smallest useful boundary.

  • Keep the local Web UI on its default 127.0.0.1 loopback address. Do not expose it to a LAN or the public internet for a first test.
  • Start in a disposable repository or worktree. The agent can read and edit workspace files, run commands, delegate work, and maintain a plan.
  • Use the Harness credential store or an environment-variable reference. Never paste an API key into a prompt, repository file, screenshot, or committed plugin configuration.
  • Treat every third-party plugin as local executable code. Review its source and dependencies before allowing it near repositories or credentials.

Configure DeepSeek Harness

1.Start the official Web UI from your project directory

The current preview serves its local Web UI at http://127.0.0.1:3080 by default.

cd /path/to/your/project
npx @deepseek-ai/dsh web
2.Add DeepSeek-V4.io under Settings → Models

Provider IDs are permanent in saved sessions. Use a lowercase stable ID and keep the model text-only. Harness stores the key as a write-only credential reference, and saved model changes apply on the next request without a restart.

Provider type: Custom provider
Provider ID: deepseek-v4-io
Display name: DeepSeek-V4.io
Base URL: https://deepseek-v4.io/api/v1
API protocol: OpenAI Completions
API key: your-deepseek-v4-io-key
Model ID: deepseek-v4-flash
3.Optional: keep the API key in an environment variable

Launch dsh from the same environment. The settings fragment below references this variable and does not contain the literal key.

export DEEPSEEK_V4_API_KEY="your-deepseek-v4-io-key"
4.Optional: merge this provider fragment into $DSH_HOME/settings.yaml

Merge this block into the existing file instead of replacing unrelated settings. It is a transparent settings fragment, not an executable third-party plugin.

llm-pi-ai:
  providers:
    deepseek-v4-io:
      displayName: DeepSeek-V4.io
      apiKeyEnv: DEEPSEEK_V4_API_KEY
      api: openai-completions
      baseURL: https://deepseek-v4.io/api/v1
      defaultInput: [text]
      models:
        - id: deepseek-v4-flash
          name: DeepSeek V4 Flash

Verify the setup

The page remains Experimental until these checks pass against the public DeepSeek-V4.io gateway.

  1. 1Use Fetch available models and confirm deepseek-v4-flash is returned, or enter it manually.
  2. 2Choose a workspace and ask the agent to summarize one file without editing it.
  3. 3Run one read-only shell or file tool call and confirm the final response and Developer usage row both complete.

Troubleshooting

The Web UI or settings moved

DeepSeek labels Harness a developer preview and warns about compatibility-breaking changes. Recheck the official README and model-provider guide before treating this page as stable.

Fetching models returns 401

Use a paid DeepSeek-V4.io key created in /developer. A DeepSeek Platform or OpenRouter key is not valid for this gateway.

The model is missing after save

Confirm the base URL ends in /api/v1, set the protocol to OpenAI Completions, and add deepseek-v4-flash manually if model discovery is unavailable.

Environment-based settings report MISSING_CREDENTIAL

Export DEEPSEEK_V4_API_KEY in the same environment that launches dsh, then start a new process. Do not place the literal key inside settings.yaml.

An image request is rejected

This DeepSeek-V4.io route is deliberately text-only and exposes deepseek-v4-flash. DeepSeek's experimental vision model is documented for the official DeepSeek API, but is not yet available through this gateway.

Sources and support boundary

Configuration fields are checked against the tool publisher's documentation. DeepSeek-V4.io owns this gateway guide and support for its endpoint; it does not represent the tool publisher.

Ready to connect DeepSeek Harness?

Create a DeepSeek-V4.io API key, copy the configuration above, and verify one read-only tool call before using it on a real project.

Open Developer Dashboard