DeepSeek Harness — Cordis plugins, agent modes & session logs

Use when

You need a composable coding-agent runtime you can audit—swap models, tools, sandboxes, and UI as plugins, with append-only session logs.

Skip when

You mainly want a polished daily-driver editor, or you need a production-stable framework with long-term API guarantees and do not want a developer preview.

Try instead

Use an editor-first AI IDE for day-to-day shipping; read the on-site harness guide when you are still choosing between runtime and editor.

• Cursor overview
• Harness plugin guide

• DeepSeek HarnessThis page

  Pricing

  Harness is free (MIT); you pay model providers and your own infra

  Learning cost

  Cordis plugins + runtime modes; developer-preview docs and CLI/Web UI

  Output limits

  Coding-agent modes, trajectory logs, swappable sandboxes/UI; APIs may change

  Privacy

  Runs locally or on your infra; model calls go to whichever providers you configure

• LangGraph

  Pricing

  Open-source core; paid LangSmith/cloud options for observability at scale

  Learning cost

  Graph/state mental model; Python/JS ecosystems and LangChain docs

  Output limits

  General agent graphs beyond coding; strong orchestration patterns

  Privacy

  Self-host or cloud; tracing may leave your network if you use hosted observability

• OpenAI Agents SDK

  Pricing

  SDK free to use; model and tooling billed through OpenAI usage

  Learning cost

  Python/TS agent loops; OpenAI docs and examples

  Output limits

  OpenAI-centric agents and tools; ecosystem tied to that provider surface

  Privacy

  Agent runs and model calls follow OpenAI platform data policies

1. DeepSeek Harness — official site
2. deepseek-ai/deepseek-harness on GitHub
3. LangGraph — official site
4. OpenAI Agents SDK — docs

Free tier

Yes

Pricing summary

DeepSeek Harness is MIT-licensed open source in developer preview—no subscription for the harness itself. You still pay whichever model providers you wire in, plus any sandbox or storage infrastructure you run.

Reviewed on 20 August 2026 · deepseek-ai/deepseek-harness on GitHub

• Everything is a plugin

  Cordis mounts and unmounts plugins with dependency awareness, so you can swap models, tools, skills, sandboxes, storage, loops, scheduling, and UI pieces without rewriting the harness core.

• Traceable session trajectory

  An append-only session log records system prompts, reasoning, tool calls and results, subagent scheduling, and context injections. Resume, fork, search, and replay operate on the same event stream.

• Standard coding agent mode

  Standard mode ships a full toolset—file editing, shell, file and web search, skills, planning, goals, subagents, and workflows—for day-to-day agent development.

• Code mode orchestration

  Code mode exposes tools through a Code Mode SDK so the model can combine multi-step operations in one TypeScript program instead of chaining single tool calls by hand.

• Minimal and Creator modes

  Minimal mode keeps only persistent bash and str_replace_editor for model benchmarking. Creator mode adds runtime inspection and in-memory Cordis plugin experiments for building custom presets.

• Open source MIT preview

  Source is available on GitHub under MIT. Core plugins and APIs may still change while DeepSeek Harness remains in developer preview.

1. 1

   Launch the Web UI with npx

   Install Node.js, then run `npx @deepseek-ai/dsh web` to open the DeepSeek Harness Web UI without cloning the repo first.

2. 2

   Clone when you need full source

   For local plugin work, clone https://github.com/deepseek-ai/deepseek-harness and follow the repository setup instructions.

3. 3

   Pick a runtime mode

   Use Standard for full coding-agent workflows, Code mode for TypeScript-orchestrated tool batches, Minimal for two-tool benchmarks, or Creator to inspect and recombine plugins.

4. 4

   Compose capabilities in configuration

   Select, swap, or extend plugins—models, tools, skills, sandboxes, storage, loops, scheduling, UI—through configuration rather than patching harness internals.

5. 5

   Inspect runs in Trajectory view

   Open the session log to audit what the model saw, verify tool results, then resume or fork from the same append-only stream when iterating on agent behavior.

• Agent harness developers

  Build or evaluate coding agents with swappable tools, sandboxes, and loops while keeping a single Cordis-based runtime.

• Researchers benchmarking models

  Minimal mode isolates shell and file-edit tools so you can compare model behavior in a constrained, reproducible environment.

• Platform and infra engineers

  Experiment with sandboxes, storage, scheduling, and UI plugins without rewriting the DeepSeek Harness core for each deployment shape.

• Teams building custom agent presets

  Creator mode supports runtime inspection and in-memory plugin trials before shipping a reusable Standard or Code mode preset.

Developer preview status

DeepSeek Harness is in developer preview; Cordis plugins and public APIs continue to evolve. Treat breaking changes as expected until a stable release.

License and source

Distributed as open source under MIT from github.com/deepseek-ai/deepseek-harness. Community plugins are listed separately from the core repository.

Quick start package

The Web UI entrypoint is `@deepseek-ai/dsh` via `npx … web`. Node.js is required; confirm current CLI flags in official developer docs.

Cordis kernel

Plugin mount/unmount and dependency wiring run through Cordis services and events so capabilities can interoperate without hard-coded glue.