DeepSeek Harness: Quick Start, Four Modes, and Plugin Boundaries
Start DeepSeek Harness in an isolated folder, choose between Standard, PTC, Minimal, and Creation, and inspect plugins before granting access.
DeepSeek Harness is an open-source environment for assembling a coding agent from replaceable parts: the model, tools, skills, session, sandbox, storage, agent loop, scheduling, and UI. DeepSeek labels it a Developer Preview, so the sensible first goal is one observable run in an isolated folder—not an immediate rollout across a production repository.
DeepSeek summarizes the design as Agent = Model + Harness. The model handles reasoning and generation; the harness supplies environmental context, tools, and the loop that continues beyond a single answer. When an agent edits the wrong file or misses command output, the cause may be the runtime configuration rather than the model alone.
If you plan to connect a model through a custom API endpoint, verify the protocol, Base URL, and authentication method before installing plugins. BetterToken.ai provides OpenAI-compatible and Anthropic-compatible interfaces for tools that accept a custom Base URL, but that does not prove compatibility with every Harness plugin. Check the provider schema and API Key setup in the BetterToken API documentation, then run a small read-only task in a test folder.
What DeepSeek released
The official DeepSeek Harness repository is MIT-licensed and built on Cordis. Cordis loads and unloads plugins and manages their dependencies; it does not supply the agent's capabilities by itself. Those capabilities come from plugins and can be selected or replaced in configuration.
This separation is useful when you want to test one variable at a time. You can compare a model change without also changing the tool layer, or inspect a session plugin without rebuilding the entire runtime. It also creates responsibility: every added plugin may change the data, permissions, and external services available to the agent.
Quick start in a disposable folder
After installing Node.js, the current official command starts the Web UI:
Use this sequence for the first session:
- Create an empty test folder or a clean clone that contains no secrets and no uncommitted work.
- Start the Web UI and choose Standard mode.
- Ask for a small read-only task, such as locating the entry point or explaining the repository structure.
- Compare the answer with the actual files and inspect Trajectory before allowing edits.
The isolation and manual comparison are general engineering controls, not a promise that Harness makes every shell command safe.
Choosing among the four modes
Standard is the practical default for a first real task. PTC is useful when the orchestration itself needs to be readable and reproducible; it does not automatically improve the result. Minimal is a diagnostic baseline, not simply a lighter everyday Standard. Creation is for people who already understand the current configuration and want to build a preset.
Use Trajectory as an audit trail
DeepSeek says the model-visible context and execution events are written to an append-only session log: system prompts, tool calls and results, subagent scheduling, and context injections. Trajectory lets you inspect those events by source, while recovery, branching, search, and replay share the same event stream.
After the first task, answer four questions from the log:
- Which files and command outputs did the agent actually see?
- Did unexpected context or a tool call enter the run?
- Did the run stop because of permissions, missing tools, instructions, or context?
- Can the successful path be reproduced without starting from a blank conversation?
If the claimed file path or command output does not match reality, fix access and context before rewriting the prompt.
Review plugins before installation
For each plugin, record:
- the missing capability it provides;
- the permissions, data, and external services it can reach;
- the signal that will reveal a regression and the rollback path.
Read source and dependencies before installing a community bundle. During Developer Preview, change one component at a time so Trajectory can show what affected the result.
Harness is worth testing when you need to separate model behavior from tools, debug a multi-step run, or create a reproducible preset. Delay production adoption if you have no disposable environment or no process for reviewing logs and diffs. Recheck the official quickstart before working with real data because compatibility-breaking changes are expected during the preview.