Early development — pre-beta

Features

Everything in DysonHarness exists for one outcome: getting large-scale work done with a coordinated team of agents instead of one overextended chat.

Orchestrate the work

A team of agents with clear roles, directed by you.

Multi-agent orchestration

Four agent modes — Work to orchestrate, Plan to design, Explore to research, and Drone to implement — each tuned for its part of the job. You direct the effort; the team carries it out.

Delegated subagents

Hand pieces of the work to subagents that run in the background and report back to the parent session. Parallel progress without losing the thread of the whole.

Plan → enhance → build → review → fix

For larger changes, the design is captured in a plan you can review before implementation begins. Fewer surprises, more deliberate results.

Persist and control

Long jobs stay on track — and under your control.

Persistent & resumable sessions

Sessions persist and resume, so ambitious work isn't lost to a restart. Come back later and pick up exactly where things stood.

Context control

Drop noisy turns from the model's context when a conversation goes sideways. Cleaner context means sharper answers on long jobs.

Per-mode tool limits

Each agent mode gets only the tools it needs. Researchers research; implementers implement — a sensible boundary that keeps work focused.

Models and providers

Strong models, your choice of where they run.

Built for capable models

Tested with GLM 5.2, GLM 4.7, Kimi K2.7 Code, and Kimi K3. DysonHarness is designed to get the most out of strong, modern models.

OpenAI-compatible providers

DysonHarness works with OpenAI-compatible providers, so you're free to choose where your models come from — cloud or local.

Managed plan providers

Compatible with managed subscription plans — including the ChatGPT Codex plan and the Kimi coding plan — so you can run DysonHarness on the plans you already pay for, no extra API keys required.

Tools and platforms

The everyday tools real work requires, in one place.

Windows browser control

On Windows, agents can drive a real browser — navigating, reading, and interacting with pages — so web-based work gets done end to end.

Workspace editing, shell & web research

Agents edit files in your workspace, run shell commands, and research on the web — the everyday tools real work requires, in one place.

Custom MCP servers

Connect your own Model Context Protocol servers to give agents organization-specific tools — internal APIs, databases, and services — right alongside the built-in ones.

Get started