Introduction

In the DeepSeek Harness (DSH) environment, it is relatively easy to let AI execute tasks, but it is difficult to let AI run an R&D team. Common pain points include: delegation regressions (executors replace decision-makers, with the CEO personally fixing bugs); organizational loss of control (unknown identities, unstaffed roles, lack of constraints); costs as an opaque ledger (hidden billing, budget overruns); and decisions without evidence (key changes lack approval records and audit trails). The dsh-company plugin solves these problems by simulating a real company: you provide decision-making, while the company provides execution—with staffing, ledgers, approvals, and audits.

What It Is

dsh-company is a decision-driven AI software company orchestration plugin for DeepSeek Harness (DSH), maintained by TiferKing. It turns the current root session into the founder and persistent, continuable subagents into employees. Through HR governance, monetary budgets, work DAGs, tickets, approvals, and other features, it builds a software development environment with an organizational structure, financial accounting, and process controls.

Core Features

The plugin provides the following core capabilities:

  1. Decision-first and HR governance: AI drafts the company plan (name, mission, charter, budget), which is launched after human approval. Recruitment and evaluation adjustments are handled by HR and require human approval. The charter text is parsed into a structured data tree and rendered directly on the Web console.
  2. Monetary budget and audit: It uses a single authoritative ledger and supports BigInt aggregation. WAL coordinates state, usage, audit, and mailbox. Usage and the full audit log are appended independently, and events.jsonl retains a bounded display window.
  3. Workflow and tickets: Work items form a dependency DAG with acceptance conditions. Humans submit tickets through the Web console; the founder triages and dispatches them, and the associated fix work is automatically resolved after completion.
  4. Typed approvals and Web console: It provides typed human approvals. The Web console includes tabs for Overview, Organization, Product, Work, Tickets, Hiring, Audit, and Approvals.
  5. Resource scheduling and recovery: It supports scheduling by resource (adaptive execution). Employees are persistent sessions and support cold recovery and transactional retries.

Installation and Enablement

Running the packaged plugin requires Node ^22.19.0 || >=24, pnpm, and a DSH rc.2 host.

dsh plugin --profile web add /absolute/path/to/dsh-company-<version>.tgz

After installation, restart the DSH Web process and refresh the original URL. Do not start an alternative server instead.

Typical Usage

You can ask the agent to found a company. For example, specify the company’s mission, product, budget, and HR cap.

In the founding request, you can directly specify the HR provider, model, and reasoning effort. The corresponding configuration snippet is as follows:

{
  "hr_provider": "provider-id",
  "hr_model": "model-id",
  "hr_reasoning_effort": "default",
  "hr_budget": 10
}

Notes

  • Remote browser limitation: The remote browser sees a read-only degraded view.
  • Snapshot safety: Snapshots never carry attempt capabilities, execution prompts, credentials, or private work evidence.
  • License: The README text does not explicitly display the license content (it only mentions the LICENSE file); please check the source code for details.
  • Environment dependency: Before installation, ensure the Node version meets the requirements and the host is DSH rc.2.

Conclusion

dsh-company turns “managing an AI team” into “running a company that is visible, manageable, and accountable”. You can learn more details through the plugin directory or view the source code on GitHub.