Skip to main content
All posts
Guide

Is DeepSeek Harness Free? License, Costs, and Savings

DeepSeek Harness itself is free, MIT-licensed, open-source software — but running it costs whatever your configured model provider charges for API usage.

DeepSeek Harness (dsh) itself is free: it's MIT-licensed, open-source software with no subscription fee, no seat license, and no paywall on the tool or its plugin ecosystem. What isn't automatically free is running it — every session ultimately makes API calls to whichever model provider you've configured, and that provider bills you for usage under its own separate pricing. This guide separates those two things clearly, without quoting specific price numbers that would go stale.

The harness is free; the model calls are not automatically free

This is the single most important distinction to keep in mind. dsh the software costs nothing to download, run, or extend — you can read every line of its source on GitHub under the MIT license, fork it, modify it, and redistribute it. Its plugin ecosystem is the same: FindHarness indexes hundreds of community plugins, and installing them costs nothing beyond pnpm's normal package-install time.

What costs money is the model provider behind your session. dsh doesn't include any bundled model access — you configure a provider under Settings → Models and that provider's own API pricing applies. Depending on which provider you choose, your actual out-of-pocket cost could range from "pay-as-you-go DeepSeek API usage" to "$0, because you pointed it at a local model" — dsh's architecture supports both.

There's also no official dsh pricing page to check, because there's nothing to price — the project doesn't sell dsh itself in any tier. Support is community-driven rather than commercial: the official feedback channels are GitHub Discussions and a Discord server, both free to use, with GitHub Issues disabled in favor of Discussions. If you're used to open-source tools that have a "free tier" and a paid "pro" tier of the tool itself, dsh doesn't work that way — the tool has exactly one tier, and the only recurring cost is the model API usage layered on top of it.

Which providers you can point dsh at

dsh's built-in provider directory covers DeepSeek's own models natively, plus a directory of other providers you can add with an API key: Anthropic, OpenAI, Amazon Bedrock, Google Vertex AI, Azure, and Codex-native auth (some of these require provider-native authentication rather than a plain API key — Bedrock needs AWS credentials, Vertex needs an ADC project, Azure needs an api-version, and Codex uses OAuth). Beyond the built-in directory, you can add any OpenAI-compatible custom endpoint — including a self-hosted or local model server, if that's your preferred way to control cost. See Use OpenAI, Anthropic, or Any OpenAI-Compatible API with DeepSeek Harness for the exact configuration steps, including the Fetch available models button and the modelOverrides YAML needed for vision-capable custom models.

Provider pathCost structure
DeepSeek nativePay-as-you-go API usage, billed by DeepSeek
Anthropic / OpenAI / Bedrock / Vertex / Azure / Codex-nativePay-as-you-go, billed by that provider under its own pricing
Any OpenAI-compatible custom endpointWhatever that endpoint charges — including $0 if it's your own local or self-hosted model

Context caching and cost efficiency

DeepSeek's own platform documents a context-caching capability that can reduce API cost when prompt content repeats across requests within a session — a documented feature of DeepSeek's API, not something dsh adds on top. Community members discussing dsh's launch have reported strong cache hit rates and meaningfully reduced spend from this in practice, but we haven't independently benchmarked specific hit-rate or savings numbers ourselves, so treat those reports as anecdotal rather than a guaranteed outcome. If minimizing spend matters to you, check DeepSeek's own API documentation for how caching is billed and structure your session/context reuse accordingly — this applies regardless of which harness you're running the model through.

Ways to keep the bill low

A few concrete, harness-level levers, distinct from provider-side pricing choices:

  1. Use a local or self-hosted OpenAI-compatible model as a custom provider — the marginal API cost per request drops to whatever your own infrastructure costs, rather than a per-token bill to a hosted provider.
  2. Scope permission presets deliberately. dsh's default new-session preset is workspace-write with ask approval — reviewing what an agent is about to do before it burns tokens on a wrong path is a real cost lever, not just a safety one.
  3. Use headless mode for scripted, bounded tasks (dsh --profile headless "task") instead of open-ended interactive sessions, where a task's scope — and therefore its token usage — is naturally more contained.
  4. Track spend with a community plugin. deepseek-harness-wallet surfaces official DeepSeek balance, per-session cost and tokens, and low-balance alerts directly in the Web UI — useful if you want visibility without checking your provider's dashboard separately.
  5. Fall back between providers deliberately rather than by accident — plugins like dsh-llm-fallbacks add role-based retry and fallback strategies, which can help avoid wasted retries hitting a rate-limited or expensive provider repeatedly.

What "free" doesn't cover

Being MIT-licensed and free to run doesn't mean there's no cost anywhere in the stack. Worth being clear-eyed about:

  • Model API usage is the main ongoing cost, as covered above, and scales with your usage pattern, not with the harness itself.
  • GitHub-sourced plugins that need a build step can require you to grant allowBuilds permission, which executes that package's install-time code on your machine — not a monetary cost, but a security cost worth weighing (see our plugin security checklist).
  • E2B cloud sandboxes, one of dsh's supported sandbox backends alongside local Linux/macOS/Windows options, are a third-party cloud service with its own separate pricing if you choose to use it instead of local execution.
  • dsh itself carries no license fee, but it's a developer preview with no SemVer guarantee — factor in the engineering cost of tracking breaking changes if you're relying on it for something production-critical.

FAQ

Is DeepSeek Harness completely free to download and use?

Yes, the software itself is MIT-licensed, open-source, and free — no subscription, no seat license, no paywalled features. Running a session still requires a configured model provider, which bills separately.

Do I have to pay DeepSeek specifically, or can I use a different provider?

You can use DeepSeek's own models, a built-in directory of other providers (Anthropic, OpenAI, Bedrock, Vertex, Azure, Codex-native auth), or any OpenAI-compatible custom endpoint — including a free local model if you're running one yourself. dsh doesn't require you to pay DeepSeek specifically.

Are dsh's plugins free too?

Yes — installing a plugin is a normal pnpm package install with no separate licensing fee that we're aware of across the plugins indexed on FindHarness. Always check an individual plugin's own license and source before installing, the same way you would for any open-source dependency.

Does context caching actually make DeepSeek cheaper?

DeepSeek's platform documents a context-caching feature that reduces cost for repeated prompt content, and community reports describe meaningful savings from it — but we haven't independently verified specific numbers, so check DeepSeek's own billing documentation rather than relying on anecdotal figures.

What's the cheapest way to run dsh?

Point it at a local or self-hosted OpenAI-compatible model as a custom provider — your API cost becomes whatever your own infrastructure costs rather than a hosted provider's per-token pricing. See Use OpenAI, Anthropic, or Any OpenAI-Compatible API with DeepSeek Harness for setup.

Next steps