DeepSeek API Docs Mirror

Unofficial markdown mirror of api-docs.deepseek.com. 134 pages (en + zh-cn), built for LLM agents: plain markdown, stable paths, llms.txt index, llms-full.txt single-file dump. Includes the FAQ, which is not part of the docs site and is not readable as text anywhere else, and the dsh (DeepSeek Harness) docs, which live on GitHub/npm rather than the docs site.

Clone this repo and point your coding agent at it. Every DeepSeek API doc -- quick start, API reference, guides, samples, agent integrations, news -- searchable, version-controlled, and offline.

Install

git clone https://github.com/thevibeworks/deepseek-docs
cd deepseek-docs

Then ask your agent anything:

claude "how do I enable thinking mode on deepseek-v4-pro?"
claude "what's the Anthropic-format base URL and which params map to what?"
claude "show me the FIM completion request schema"

The FAQ

content/*/faq/ does not come from the docs site. DeepSeek's FAQ lives at static.deepseek.com/faq/, whose index.html is a 562-byte shell — the content is a JSON.parse('...') blob inside a content-hashed JavaScript chunk, held as mdast rather than HTML. scripts/faq.py finds the current chunk, decodes the blob and renders the AST back to markdown, so the result round-trips rather than being scraped out of rendered DOM.

It is worth the trouble: 44 questions per locale, 15 of them about the API, covering things api-docs never mentions — invoices, refunds, what to do about a leaked key, how to request a higher rate limit.

uv run scripts/faq.py            # refresh content/<locale>/faq/
uv run scripts/faq.py --check    # exit 1 if stale

dsh (DeepSeek Harness)

content/en/dsh/ mirrors deepseek-ai/deepseek-harness (launched 2026-08-13) -- the repo README, its English docs/ tree, and a metadata page for the @deepseek-ai/dsh npm package. These docs live on GitHub/npm, not on api-docs, so the main fetcher never sees them.

dsh is a developer preview that ships breaking changes without migration paths (the .dsh-plugin manifest format and dsh-plugin-prepare were deleted 2026-08-09, cold). That is exactly why it is mirrored here: the sync diff is the changelog upstream does not write. Pages deleted upstream are pruned, so deletions show up in the diff too.

Community companions to this mirror: howto-dsh (verified traps, dated against specific dsh versions), awesome-dsh-plugins (spam-filtered plugin registry), and the dshworks org index.

uv run scripts/dsh.py            # refresh content/en/dsh/
uv run scripts/dsh.py --check    # exit 1 if stale

Content

content/
  en/                       English docs (63 pages)
    index.md                Your First API Call
    quick_start/            pricing, rate limits, error codes, token usage
    quick_start/agent_integrations/   Claude Code, Codex, OpenCode, ... (18 tools)
    api/                    API reference: chat completions, responses,
                            FIM, models, balance (from the OpenAPI-rendered pages)
    guides/                 thinking mode, tool calls, JSON mode, KV cache,
                            Anthropic API, Responses API, FIM, prefix completion
    api_samples/            curl / python / nodejs request samples
    news/                   release notes (V3, R1, V3.1, V4, ...)
    faq/                    the FAQ, one file per category, extracted from
                            static.deepseek.com (see above)
    dsh/                    dsh (DeepSeek Harness) docs from GitHub + npm
                            (see above); index.md = repo README, docs/ =
                            the repo's English docs tree, npm.md = package
    faq.md  updates.md  prompt-library.md
  zh-cn/                    Chinese mirror, identical paths (api-docs only)
llms.txt                    curated index of every English page
llms-full.txt               all English api-docs pages concatenated (~330 KB;
                            the dsh section is indexed but not concatenated)
sources.json                machine-readable source registry

Every file carries frontmatter with its canonical source: URL and fetch date. Internal links are rewritten to relative .md paths so they resolve inside the repo; images point back at the live site.

Fetching

# Requires: uv (https://docs.astral.sh/uv/)
uv run scripts/fetcher.py                # fetch everything (en + zh-cn)
uv run scripts/fetcher.py --locale en    # one locale
uv run scripts/fetcher.py --incremental  # skip existing files
uv run scripts/fetcher.py --index-only   # rebuild llms.txt / llms-full.txt
uv run scripts/fetcher.py --tree         # show sources + counts

The site is Docusaurus v3 with no llms.txt and no markdown variants, so the fetcher discovers pages via sitemap.xml (per locale) and converts the server-rendered HTML article body to markdown locally. The prompt library is the one client-rendered page; its data is fetched from the site's static prompts.json instead.

Automation

fetch-deepseek-docs.yml runs every 6 hours: fetch (dsh from GitHub/npm, then the api-docs site), then hand the diff to Claude Code running on the DeepSeek API itself (deepseek-v4-flash via the Anthropic-compatible endpoint, anthropic_api_key = a DEEPSEEK_API_KEY repo secret). The agent only classifies the change and writes decision files; a deterministic bash step publishes -- minor changes commit straight to main, high-signal changes (new models, pricing, API schema, dsh breaking changes) open a PR for human review. A deterministic pricing tripwire overrides the agent: any pricing.md diff touching time-of-day tier language (时段 / off-peak / peak / discount / multiplier / 分时) is forced high-signal, because the peak-hour plan was dropped 2026-08-08 with a repricing pending and downstream consumers need to hear about its return the same day. DeepSeek docs, kept fresh by DeepSeek.

Disclaimer

Unofficial mirror for educational and development purposes. Documentation content belongs to DeepSeek; for official docs visit api-docs.deepseek.com. Code in this repo is MIT-licensed.