DeepSeek Harness setup

Connect DeepSeek Harness to UncensoredLabs using a custom OpenAI Chat Completions provider.

Configure the client

Open Settings → Models → Add model provider → Custom model API. Enter these values, save the provider, and select its model in a new session.

Connection settings

Provider ID: uncensoredlabs
API protocol: OpenAI Chat Completions (openai-completions)
Base URL: https://uncensoredlabs.dev/v1
API key: paste your application API key into the credential field
Model ID: deepseek-v4.1-flash-1017-abliterated
Context window: 1000000
Maximum output: 65536
Input types: text

For profile-based configuration, merge this provider into the existing llm-pi-ai configuration. Preserve other providers. Make the API key available in the environment that starts Harness.

cordis.patch.yml

- id: llm-pi-ai
  config:
    providers:
      uncensoredlabs:
        apiKeyEnv: UNCENSOREDLABS_API_KEY
        api: openai-completions
        baseURL: https://uncensoredlabs.dev/v1
        models:
          - id: deepseek-v4.1-flash-1017-abliterated
            input: [text]

Leave reasoning-level and image-capability overrides off for the first connection test. Enable only settings supported by the selected model and gateway.

Official client documentationProvider configuration and version-specific options.

Verify the connection

Use an application API key from the dashboard, not your account sign-in key. Keep credentials outside committed configuration.

Start a new session and request a short response. Then ask the client to list files without changing them, and review its tool permissions before a longer task.

Configuration examples were checked against the linked official documentation on November 1, 2026. Client versions and features vary; these guides do not certify every client release.

Troubleshooting

Use text Chat Completions. Responses, Anthropic Messages and image input are not supported. Client web-search features may require a separate service.

Status	What to check
400 / 404	Check the base URL, exact model ID and Chat Completions mode. Do not append the completions path to a base-URL field.
401 / 403	Check the application API key, its model permissions and whether it has been revoked.
402	Check available credit and existing reservations. A reservation can exceed the final charge; use a smaller output limit before retrying.
429	Respect Retry-After and reduce concurrent tasks. Do not immediately retry the same request in a loop.