DeepSeek

DeepSeek provides powerful AI models with an OpenAI-compatible API.

Install plugin

Install the official plugin:

openclaw plugins install @openclaw/deepseek-provider

Installation applies to a running Gateway automatically; otherwise it takes effect on the next startup. See Apply changes and inspect.

Getting started

openclaw onboard --auth-choice deepseek-api-key

openclaw models list --provider deepseek

openclaw models list --all --provider deepseek

Onboarding preserves your model entries and leaves generated catalog rows to discovery. With models.mode: "replace", it also writes the built-in catalog because that mode skips discovery.

openclaw onboard --non-interactive \  --mode local \  --auth-choice deepseek-api-key \  --deepseek-api-key "$DEEPSEEK_API_KEY" \  --skip-health \  --accept-risk

Built-in catalog

OpenClaw's local costs are estimates. deepseek/deepseek-flash uses DeepSeek's peak rates: $0.30 per million input tokens, $1.20 per million output tokens, and $0.006 per million cached input tokens. Published off-peak rates are half those amounts. Legacy rows retain their earlier bundled metadata. DeepSeek still accepts deepseek-v4-flash and deepseek-v4-flash-vision-exp, routes them to V4.1 Flash, and bills them at current Flash rates. Existing explicit selections remain valid in OpenClaw. DeepSeek can change rates; its Models & Pricing page is authoritative for billing.

For image inputs, select deepseek/deepseek-flash. The legacy deepseek/deepseek-v4-flash-vision-exp selection also retains image support. deepseek/deepseek-flash accepts PNG, JPEG, GIF, and WebP images through the same API and API key. See DeepSeek vision for image limits.

Thinking and tools

When a DeepSeek model emits DSML tool markup as text, including through OpenRouter or another OpenAI-compatible provider, OpenClaw recovers valid calls, tool_calls, tool_call, and function_calls blocks into tool calls instead of showing their payloads in chat. This includes doubled fullwidth markers with a separator space. Direct provider calls and tool-free finalization hide these blocks without recovering tools. Malformed or unfinished blocks are also hidden without executing them; existing tool permissions still apply to recovered calls. Native tool calls take precedence when the same response also contains DSML, so recovered duplicates do not execute. Routing through OpenRouter keeps OpenRouter's reasoning request format.

DeepSeek V4 thinking sessions require replayed assistant messages from a thinking-enabled turn to include reasoning_content on follow-up requests. OpenClaw's DeepSeek plugin backfills that field automatically, so normal multi-turn tool use works on deepseek/deepseek-flash, deepseek/deepseek-v4-flash, deepseek/deepseek-v4-flash-vision-exp, and deepseek/deepseek-v4-pro even when history came from another OpenAI-compatible provider (no native reasoning_content) or from a plain assistant message. No /new required after switching providers mid-session.

When thinking is disabled (including the UI None selection), OpenClaw sends thinking: { type: "disabled" } and strips replayed reasoning_content from outgoing history, keeping the session on the non-thinking DeepSeek path.

Fresh onboarding selects deepseek/deepseek-v4-pro. To select Flash:

openclaw models set deepseek/deepseek-flash

Live testing

To run only the DeepSeek V4 direct-model checks from the modern model live suite:

OPENCLAW_LIVE_PROVIDERS=deepseek \OPENCLAW_LIVE_MODELS="deepseek/deepseek-v4-flash,deepseek/deepseek-v4-pro" \pnpm test:live src/agents/models.profiles.live.test.ts

Verifies both V4 models complete and that thinking/tool follow-up turns preserve the replay payload DeepSeek requires.

To check the experimental vision model with the same DEEPSEEK_API_KEY:

OPENCLAW_LIVE_DEEPSEEK_MODEL=deepseek-v4-flash-vision-exp \pnpm test:live extensions/deepseek/deepseek.live.test.ts

This runs text, generated-image recognition, and thinking replay checks against the selected model.

Config example

{  env: { vars: { DEEPSEEK_API_KEY: "sk-..." } },  agents: {    defaults: {      model: { primary: "deepseek/deepseek-v4-pro" },    },  },}

Related

Web search

The same plugin can provide web_search for any model. Search uses DeepSeek's search engine directly, not a chat completion. Install the plugin as above, then choose DeepSeek in Settings → Search or openclaw configure --section web.

{  tools: {    web: {      search: { provider: "deepseek" },    },  },  plugins: {    entries: {      deepseek: {        config: {          webSearch: {            apiKey: "YOUR_DEEPSEEK_API_KEY", // pragma: allowlist secret          },        },      },    },  },}

The key is optional when DEEPSEEK_API_KEY is set in the Gateway environment or models.providers.deepseek.apiKey is configured for DeepSeek's own API. Credential precedence is plugin search key, environment key, then model-provider key. Search treats DEEPSEEK_API_KEY as a DeepSeek key. It reuses the model-provider key only while models.providers.deepseek.baseUrl and every model baseUrl under it are unset or use https://api.deepseek.com (any path); a key configured for another base URL is not sent to DeepSeek and does not make DeepSeek eligible for auto-detection. Only routes in models.providers are checked, not routes that come only from an agent's models.json. The plugin search key and model-provider fallback support SecretRefs through the standard secret preparation flow. Search does not reuse API keys stored only in model auth profiles.

With DeepSeek selected as the search provider, changing its search key or SecretRef hot-applies through the Gateway's secrets runtime without replacing the plugin or waiting for unrelated agent turns. Plugin enablement and explicit plugin reloads still use normal lifecycle settlement.

With no explicit search provider, credential auto-detection considers DeepSeek after Parallel (order 80), before configured endpoint providers. Select provider: "deepseek" to prevent automatic fallback to another search service.

Search always calls https://api.deepseek.com/v1/tools/search; a configured chat base URL does not redirect search to a proxy. When chat uses a proxy, store the proxy key in a DeepSeek auth profile or in models.providers.deepseek.apiKey with auth: "api-key" (otherwise DeepSeek auth profiles and DEEPSEEK_API_KEY take precedence over the config key), and give search a DeepSeek key through the plugin search key or DEEPSEEK_API_KEY. A proxy that forwards a DeepSeek key also needs that key set for search.

Queries and results

• query: 1–128 Unicode characters. Oversized queries are rejected, never truncated.
• count: 1–10 results, defaulting to tools.web.search.maxResults or 5.
• Use query operators such as site:, -site:, lang:, after:, before:, and quoted phrases directly in query. Separate structured filter parameters are not supported and return an error rather than being ignored.
• Results contain titles, URLs, model-oriented summaries (falling back to short snippets), publication dates, and site names. Core validates URLs/dates and wraps returned prose as untrusted web content. Full page content is not requested.

await web_search({  query: "OpenClaw site:github.com lang:en after:2026-10-01",  count: 5,});

Search uses the shared timeout and cache settings. Cached results are separated by query, result count, and credential. Errors are not cached or automatically retried by the DeepSeek adapter.

DeepSeek documents a limit of 5 search requests per second per account and a price of $0.75 per 1,000 successful non-empty search requests. Empty results and failed requests are not billed. See DeepSeek Search & Fetch for current pricing and account limits. This plugin integration adds search only; web_fetch continues to use its existing local extraction and configured fallbacks.