ZeroClaw + DeepSeek

Running an AI assistant on your own hardware is only "cheap" if the model bill is cheap too — which is why the natural pairing for a self-hosted agent runtime is DeepSeek, one of the lowest-priced frontier-grade APIs you can point it at. We've already compared ZeroClaw vs OpenClaw as runtimes; this piece digs into the half nobody has written up: does ZeroClaw actually work with DeepSeek's current API, as of late September 2026?

What the integration officially looks like

ZeroClaw is provider-agnostic by design: its README names Anthropic, OpenAI, and Ollama, "and ~20 others," with fallback chains and routing for when a provider flakes. DeepSeek is in that list. Configuration lives in one TOML file at ~/.zeroclaw/config.toml, where you define a provider entry under the universal [providers.models.<type>.<alias>] schema and point an agent at it. Because DeepSeek exposes an OpenAI-compatible API, it slots into the OpenAI-compatible provider type — the same three-minute setup we documented for calling DeepSeek V4 with the OpenAI SDK maps onto ZeroClaw's config with different field names (the field-level details belong to ZeroClaw's own provider configuration docs).

The V4 compatibility saga, on the record

This is the part most quick-start posts skip, and it's the reason "just point ZeroClaw at DeepSeek" used to fail for people:

April 24, 2026 — a user files issue #6059: ZeroClaw is incompatible with DeepSeek's V4 API format. The confirmed root cause, per the fix discussion: DeepSeek V4 requires reasoning_content to be passed back in the conversation history on multi-turn requests. If your runtime drops or mangles that field between turns, the API rejects the conversation.

May–June 2026 — the project merges a chain of fixes, each closing one leak: #6107 captures reasoning_content from streaming responses, #6285 preserves it through context compression, #6284 fixes plain-text assistant turns (branch name fix/6233), and #6980 preserves it in native tool requests. Issue #6059 itself was closed June 2, 2026.

As of September 22, 2026 — a title-level sweep of the ZeroClaw issue tracker turns up no open DeepSeek-specific compatibility issues. The nearest living relative is #5600, an open issue about preserving reasoning_content for Kimi Code — same subsystem, different provider. That's the honest state of "is V4 compat a continuous engineering effort": it was, the DeepSeek branch of it is closed, and its Kimi sibling is not.

The model-name trap (this is the one that will bite you)

DeepSeek's own changelog, dated September 10, 2026: V4.1 Flash is live and you call it with the model name deepseek-flash. The previous generation — V4 Flash and V4 Flash Vision — is retired, and the old names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1 Flash for compatibility. "Temporarily" is the operative word: that alias is a courtesy, not a contract. Meanwhile DeepSeek V4 Pro remains available past its originally announced September 14 retirement date, billing unchanged (the retirement was canceled), so deepseek-v4-pro configs keep working — for now, with DeepSeek explicitly reserving the right to change that.

Practical takeaway for your config.toml: write deepseek-flash for the current cheap multimodal model, or deepseek-v4-pro if you specifically need the larger one. If your ZeroClaw install predates September 10, double-check which model string your quickstart wrote — a retired-but-routed name will work today and fail without warning on the day DeepSeek pulls the alias. We cover the pricing side of the same reshuffle in the V4.1 Flash launch note.

What the combo costs

The whole point of self-hosting the runtime is that the model bill becomes your only real cost — and DeepSeek is the aggressive end of that market. Between off-peak discounts and automatic context caching (cache hits bill at a fraction of the input rate — a meaningful lever for an always-on agent that re-reads long histories every turn), the same agent workload can vary by multiples depending on how you schedule and what your prompt structure reuses. Our cheapest DeepSeek V4 API providers breakdown covers the hosting alternatives if you'd rather not use the first-party API at all.

ZeroClaw + DeepSeek vs OpenClaw + DeepSeek

For balance: OpenClaw's DeepSeek story is more formally blessed — DeepSeek's own API documentation ships an OpenClaw integration page, which is as first-party as third-party-tool support gets. ZeroClaw's support is equal in kind (it's in the provider list) but has the bumpier recent history documented above. Neither runtime charges you anything itself; the differentiation is footprint versus ecosystem, which we laid out in the ZeroClaw vs OpenClaw comparison.

If your agent backend is DeepSeek and your hardware is a $5 VPS or a Raspberry Pi, ZeroClaw on v0.8.5 with deepseek-flash configured is a defensible, current, cheap stack — with the caveat that this page reflects the official trackers as of the date below, not a lab test.

Checked September 22, 2026. ZeroClaw facts verified against the official repository (v0.8.5), its issue tracker (#6059, #6233, #5600, PRs #6107/#6284/#6285/#6980), and docs.zeroclaw.com; model names and retirement dates against DeepSeek's official changelog (api-docs.deepseek.com/updates). This article is based on official sources and issue trackers — no hands-on testing was performed. Issue states and model names reflect the checked date and can change.

Sources: ZeroClaw official repository · Issue #6059 — Incompatible with DeepSeek-V4 API format · PR #6284 — preserve reasoning_content for plain-text turns · Issue #5600 — Kimi Code reasoning_content (open) · DeepSeek official Change Log · ZeroClaw Docs