Categories Alternatives News Submit a Tool Advertise About

What Does DeepSeek Harness Cost to Run?

The runtime is $0 and MIT-licensed — there is no paid tier behind it. Your real bill is the model: on DeepSeek's API, Flash is $0.15 per M input (cache miss) and $0.60 per M output off-peak, Pro costs 4.4× that, and every token is half price outside peak hours — including all weekend.

MIT222K+ Stars$0 runtimeBYO model
7.8 / 10
AICoding score
Visit Official Site ↗
Quick answer
$0 MIT runtime, self-hosted — no subscription or paid tier exists
$0.15 / $0.60 Flash per M tokens off-peak: cache-miss input / output
50% Off-peak discount — weekends are entirely off-peak
4.4× What V4-Pro costs per token versus Flash
Updated November 1, 2026 Based on research & community feedback
License MIT (open source)
Runtime cost $0 — no paid tier
Core Cordis plugin kernel
Models Flash $0.15/M in, Pro $0.66/M in (off-peak)
Off-peak 50% of peak — weekends all off-peak
Built-in tools 53
Plugins 700+ community
Status Developer preview (breaking changes expected)

Quick Verdict

7.8 / 10 — the most extensible agent runtime in the category and genuinely free software, held back only by developer-preview maturity and an ecosystem that is still stabilising.

dsh makes the entire agent loop replaceable: model adapter, tools, sandbox and UI are all plugins over the Cordis kernel. Because there is no vendor to pay, your cost is entirely the model bill — which makes the cheap, cache-friendly and off-peak choices on this page the ones that decide what you actually spend.

Use DeepSeek Harness if…

  • You want to own the agent runtime and pay no vendor at all
  • You are happy to tune cost yourself: Flash versus Pro, off-peak runs, cache-friendly prompts
  • You need the append-only, replayable session log for audit or debugging
  • You are building an agent product on top of a plugin kernel rather than forking a closed tool

Skip DeepSeek Harness if…

  • You need production stability this quarter — the repo is an explicit developer preview
  • You want a hosted product with support, SLAs and no infrastructure work
  • You will not read SAFETY.md before giving an agent shell and filesystem access
  • You need first-class vision from the model itself — DeepSeek V4-Pro cannot see images

Core Features

The defining idea

Everything Is a Plugin

The core design principle and the reason the ecosystem exploded. The model adapter, tool registry, session log, sandbox, approval policy and even the UI are all plugins over the Cordis kernel — no privileged core, extensions mount as siblings and unwind on unload. You replace any capability from a config file without forking source.

Ready out of the box

53 Built-in Tools

bash, pwsh, edit, read, write, glob, grep (bundled ripgrep), lsp (language server protocol), subagent + subagent_fork, web_fetch, persistent terminal sessions, and the cordis_* family. The catalog is generated by actually booting each plugin and reading its schemas, with a build guard that fails if a tool package is missing.

Close the vision gap

Vision & Image Recognition via Plugins

DeepSeek's V4 models are text-only, but ModLens and dsh-vision-toolkit convert images into structured evidence (text, layout, entities) before reasoning, and open-design (87k+ stars) turns the harness into a local design workbench that generates and previews prototypes, webpages, slides, images and video. Vision is one plugin away, not a model limitation.

Every run is traceable

Traceable by Design

An append-only session log records everything the model saw — system prompts, reasoning, tool calls, subagent scheduling, raw streaming chunks — replayable down to the failed step. The invariant is "model-visible means logged". For debugging agent runs and compliance auditing, this is the strongest replay story of the harness category.

Beyond single agents

Multi-Agent Teams & Orchestration

dsh-agent-teams promotes an agent to team leader, splits tasks, sets dependencies, and lets members message each other, with the Web UI showing real-time status of who is working. Combined with the 4 runtime modes (Standard, Code/PTC, Minimal, Creator), it is a full orchestration surface.

Skills are portable

Migration & Long-Term Memory

dsh-plugin-claude-bridge imports Claude memories, skills and CLAUDE.md; dsh-claude-move migrates whole sessions so you continue the conversation in Harness. dsh-memory-evolve adds cross-session memory that persists project conventions and decisions across sessions — turning the agent from stateless chat into a project-aware assistant.

Pricing

DeepSeek Harness
$0
free forever, MIT

The whole runtime — plugin kernel, 53 built-in tools, all runtime modes, Web UI and TUI. There is no paid tier behind the free one: no hosted plan, no enterprise SKU, nothing to upgrade to.

  • MIT open source
  • All runtime modes
  • Web UI + TUI
  • No paid tier exists
Flash (V4.1-Flash)
$0.15 / $0.60
per M tokens off-peak

Cache-miss input at $0.15 and output at $0.60 per million tokens off-peak; cache hits cost $0.003, which is where long agent sessions get cheap. Peak hours double the price. Supports vision.

  • $0.003/M cache hits
  • Peak doubles the rate
  • Vision supported
  • 1M context, 384K output
V4-Pro (0813)
$0.66 / $1.98
per M tokens off-peak

Roughly 4.4x Flash per token for harder reasoning: $0.66 cache-miss input and $1.98 output per million off-peak, $1.32 and $3.96 at peak. No vision. Use it deliberately rather than by default.

  • 4.4x Flash per token
  • No vision support
  • Lower concurrency (500)
  • Best for hard tasks only
Local weights
$0
your hardware

Run open-weight models through Ollama or an OpenAI-compatible endpoint: no token bill at all, and code never leaves your machines. You pay in GPU time and slower runs instead.

  • $0 tokens
  • Fully on-prem
  • Slower and weaker
  • Needs a decent GPU

* Model prices verified against DeepSeek's official API pricing page on September 13, 2026. Note the shape of the discount: off-peak is exactly half of peak, peak is only 01:00-04:00 and 06:00-10:00 UTC Monday to Friday, so weekends are billed entirely off-peak. DeepSeek reserves the right to change prices.

What DeepSeek Harness Really Costs

A free runtime, and the three token levers that decide your bill
Task / Plan Usage Typical cost What to watch
The software Self-hosted runtime, all plugins $0 MIT — no hosted or enterprise tier exists
Flash, off-peak Routine agent work outside peak hours $0.15 in / $0.60 out per M The default choice, and where most projects should start
Flash, cache hits Long sessions with stable prompts $0.003 per M 50x cheaper than a cache miss — the single biggest lever
Flash, peak hours UTC 01:00-04:00 and 06:00-10:00, Mon-Fri $0.30 in / $1.20 out per M Exactly double: schedule heavy runs around it
Weekend runs Batch jobs, migrations, test sweeps Off-peak all day The cheapest window of the week
V4-Pro, off-peak Hard reasoning, large refactors $0.66 in / $1.98 out per M 4.4x Flash — use it deliberately, not by default
V4-Pro, peak Same work during peak hours $1.32 in / $3.96 out per M The most expensive configuration on this page
Local weights Ollama or on-prem GPU $0 tokens You pay in hardware and slower iterations
Concurrency Parallel agent runs No charge, but capped Flash 2,500 vs Pro 500 concurrent requests
Vision work Screenshots, UI reasoning Forces Flash V4-Pro cannot see images, so vision pins you to the cheap model

Two facts shape every bill here: cache hits cost 50x less than cache misses ($0.003 vs $0.15 per million on Flash), and off-peak is exactly half of peak with weekends entirely off-peak. Model choice matters more still — V4-Pro costs 4.4x Flash per token. Figures verified September 13, 2026 against DeepSeek's API pricing page; token spend at your own usage level is yours to measure.

DeepSeek Harness Deals & Ways to Spend Less

Offers, free tiers & ways to pay less
Free Free open source — MIT license, no paid tier exists Details →
Money-Saving Run heavy work off-peak or at weekends — DeepSeek API bills at 50% Details →

There is nothing to discount — the software is already free. The real savings are cache hits, the Flash model, and running heavy work off-peak or at the weekend.

vs The Field

dsh vs Claude Code

Open loop vs closed vertical

Claude Code is a closed vertical: Anthropic's loop around Anthropic's models. dsh is the opposite — MIT-licensed, plugin-driven, model-agnostic, and it can even migrate your Claude sessions (dsh-plugin-claude-bridge). Claude Code wins on maturity and first-party support; dsh wins on freedom, cost control and no lock-in. Most teams keep both: Claude Code for production, dsh to explore.

dsh for openness · Claude Code for maturity

dsh vs Codex CLI

Both open source, opposite architecture

Codex CLI is Apache-2.0 but architecturally a closed vertical around OpenAI — the loop is fixed, extensibility is via plugins on OpenAI's terms. dsh makes the entire loop replaceable and is model-agnostic. Codex is more stable and better integrated with ChatGPT plans; dsh is the assembly kit. For developers who want to own the runtime: dsh.

Codex for stability · dsh for ownership

dsh vs OpenHands

Open agent framework, different focus

OpenHands is a turnkey open-source agent you run, with a UI and workflow out of the box. dsh is a plugin meta-framework you assemble — more setup, far more flexibility, and a much faster-growing plugin ecosystem (700+ vs OpenHands' extension model). OpenHands if you want working software today; dsh if you want to build your own agent platform.

OpenHands for turnkey · dsh for customization

What It Is Worth Paying For

Escape vendor lock-in

MIT license + plugin architecture + model-agnostic adapters mean no single vendor owns your agent loop. Swap models, sandboxes and UI without changing the harness — the anti-closed-vertical choice.

Cheapest path:Free + BYO key

Building an agent product

Use dsh as a composable base instead of forking Claude Code or wrapping Codex. Cordis dependency injection and reversible effects make plugins the extension model for your product.

Cheapest path:Free + plugins

Compliance & audit

The append-only, token-level-replay session log is the strongest audit story in the category — model-visible means logged, so every agent action is reconstructible.

Cheapest path:Free (self-hosted)

Multi-agent orchestration

dsh-agent-teams turns the harness into a team manager: leader splits tasks, subagents message each other, and the Web UI shows live status. Built for the 2026 multi-agent standard.

Cheapest path:Free + agent-teams

Claude Code migration

If you are leaving Claude Code, dsh-plugin-claude-bridge and dsh-claude-move import skills, memory and sessions so you pick up where you left off — no rewrite.

Cheapest path:Free + bridge plugin

Cheapest Way to Run It

Three decisions that set the token bill: launch, model choice, and when you run
01

Launch it with one command

Run npx @deepseek-ai/dsh web and the Web UI starts on http://127.0.0.1:3080 — no install script, no account, no vendor signup. That is the whole software cost of the project.

02

Choose the model as your cost dial

Flash for routine work, Pro only for genuinely hard reasoning at 4.4x the price, or local weights for $0 tokens. Keep prompts stable: cache-hit input costs $0.003 per million instead of $0.15, which is the biggest single saving available.

03

Schedule heavy work off-peak

Peak is only 01:00-04:00 and 06:00-10:00 UTC Monday to Friday. Everything else — every night, every weekend — bills at half price. Batch migrations and test sweeps accordingly.

FAQ

Is DeepSeek Harness free?

Yes — fully free and MIT licensed. The harness itself costs $0 forever, on every machine. You only pay for model API usage (or run local models via Ollama for $0). "Everything is a plugin" means even the UI is replaceable, so there is no paid tier hiding behind the free one — the project is MIT open source with no commercial licensing gate.

Is DeepSeek Harness a model?

No. dsh is an agent runtime (a harness) — the layer that runs models. It is DeepSeek's answer to Claude Code and Codex CLI, but instead of bundling a model, it lets you plug in any model: DeepSeek, Claude, GPT, Gemini, or a local Ollama model. "Everything is a plugin" applies to models too — swap the adapter from a config file without changing the harness.

How is DeepSeek Harness different from Claude Code / Codex?

Three structural differences. (1) It is open source (MIT) and the agent loop itself is open — Claude Code and Codex are closed verticals around their own models. (2) Everything is a plugin: model adapter, tools, session log, sandbox, even the UI are swappable plugins over the Cordis kernel. (3) It is a kit, not a turnkey product — more setup work, far more freedom. Same category, opposite design philosophy.

How many DeepSeek Harness plugins are there?

Over 700 public repos carry the dsh-plugin topic (within 24 hours of release), and the index site dsh-index.xlings.org aggregates them into a searchable catalog. They range from productivity (multi-agent teams, cross-session memory, Claude Code migration, vision) to themes and games. The community compares the ecosystem to Minecraft: dsh is the survival mode, plugins are mods, awesome-dsh-plugins is CurseForge.

Can DeepSeek Harness see images / do vision?

Not natively for the DeepSeek V4 family (text-only), but the plugin model fixes it: ModLens and dsh-vision-toolkit extract text, layout and entities from images into structured evidence before reasoning, and open-design (a local design workbench) lets the agent generate and preview prototypes, webpages, slides, images and video. Vision is one plugin away rather than a model limitation.

Can I migrate from Claude Code to DeepSeek Harness?

Yes — two plugins handle it. dsh-plugin-claude-bridge imports Claude memories, skills and configuration files; dsh-claude-move performs a full migration of sessions, memory, skills and CLAUDE.md so you continue the same conversation inside Harness. This mirrors the wider 2026 trend of skills becoming portable across agent ecosystems.

Is DeepSeek Harness stable for production?

Not this month. It is a v0.1 developer preview that explicitly warns of compatibility-breaking changes. The plugin ecosystem is moving extremely fast (700+ repos in days), so pin versions and treat it as an assembly base. For production-critical work, pair it with budget models for routine tasks and keep a stable fallback tool until the preview matures.

What models can DeepSeek Harness use?

Any OpenAI-compatible API: DeepSeek (V4-Pro, V4-Flash), Claude, GPT, Gemini, or local models via Ollama. The recommended default is DeepSeek V4-Flash for budget and V4-Pro for complex work — the same routing pattern teams now use across all harnesses: cheap model for routine steps, frontier model for hard ones.

How much does it cost to run?

The tool is free; the bill is model usage. DeepSeek V4-Flash is $0.22 / $0.66 per M tokens off-peak (and weekends are now all-day off-peak), with V4-Pro at $1.32 / $3.96 and cached input at $0.44 — far below subscription pricing. Local models via Ollama cost $0. A typical daily agent workload with V4-Flash lands in low single-digit dollars per month.

Does DeepSeek Harness have a GUI?

Yes, several. The official Web UI (dsh-web-app) runs on http://127.0.0.1:3080, dsh-TUI gives a full-screen terminal with status meters, and community clients like deepseek-harness-desktop (10k+ stars) package the whole thing without requiring Node. Because UI is a plugin, you are not stuck with one interface.

How much does it cost to actually run it every day?

On DeepSeek's own API, Flash costs $0.15 per million input tokens on a cache miss and $0.60 per million output tokens off-peak — and half of a long agent session is often cache hits, which cost $0.003 per million. A routine daily agent workload therefore lands in low single-digit dollars per month. Peak hours (UTC 01:00-04:00 and 06:00-10:00, Monday to Friday) are double, and Pro costs 4.4x Flash per token ($0.66 in / $1.98 out off-peak), so the model choice and the clock are the two levers that matter more than anything the harness does.

Is there a paid or hosted version of DeepSeek Harness?

No. The repository is MIT-licensed and the maintainers publish no hosted tier, no subscription and no enterprise SKU — you run it yourself with your own model keys or local weights. That is the entire business model of the project, and it is unusual even among open-source agents: there is nothing to upgrade to, so the only recurring cost is what your model provider charges.

Is DeepSeek Harness ready for production?

Not yet, and the maintainers say so explicitly: the repository is a developer preview and warns that it will contain compatibility-breaking changes, with a SAFETY.md that should be read before granting it filesystem and shell access. There is no SLA, no vendor support desk and no stability promise. The honest use pattern is: run it in an isolated environment or container, keep it away from secrets you would not hand an experimental agent, and treat plugins — which can mount into the runtime like siblings — as you would any third-party dependency.