Source: https://harnsy.dev/features/models

> Claude Code, Codex, OpenCode and pi work in one team. Each does what it’s best at, and you’re not tied to one vendor or its limits.

# Take the best from every model.

Claude Code, Codex, OpenCode and pi work in one team. Each does what it’s best at, and you’re not tied to one vendor or its limits.

## An independent opinion.

The best teams bring in outside consultants: an insider stops seeing what they’re used to, an outsider doesn’t. Let Codex review the code Claude wrote. Or ask three models from three vendors the same question and get three independent answers.

## Routine work goes to cheap and local models.

Don’t spend your flagship’s limits on simple work. Hand data labelling, reading and sorting texts, and rough checks to a local or low-cost model through OpenCode. The flagship sets the task and gets the answer back.

## The right model for the job.

Models have different strengths. In the author’s experience, Codex writes better texts and articles, and Claude writes better code. Build a team where everyone does what they do best.

## Your own models, through OpenCode.

OpenCode runs local models with no vendor lock-in. It’s a direct way to put your own model on the team: harnsy delivers its tasks and brings back its answers like any other agent’s.

## One launch tree for everything.

Choose the AI program, the model and the effort for a machine, an organization, a project, a team or a single role; a lower level overrides a higher one. Agents already running can be moved over.

The weekly limit ran out on Wednesday. The work didn't.

## Claude can hand routine work to another model.

Howthrough OpenCode a local or low-cost model joins the team, and harnsy carries its tasks.

site-3f → qwen-local

> I train models. Claude runs the whole training process. Data labelling I gave to GLM on a subscription and to Qwen, which runs locally on my machine. They do it pretty well, and I don’t spend my Claude subscription on labelling.
> 
> Pavel Buchnev · CTO, author of harnsy

Or ask Codex to review code written by Claude.

## Limits, honestly

- Savings aren’t measured yet: the cost of a team depends on how many agents you run.
- The model comparison is the author’s observation, not a benchmark.
- OpenCode is free; the model behind it may not be.

## Your first team is one prompt away.

Installing takes one sentence to your agent. To run your agents, harnsy drives a terminal: tmux if it is installed, otherwise its own terminal host (on Windows, always its own).

`Install harnsy following https://harnsy.dev/llms.txt`

## Questions

What&#39;s a harness?

The program an AI agent runs in: Claude Code, Codex, OpenCode, pi. harnsy doesn't replace them; it connects them.

What will it cost in total?

harnsy adds no AI charges: your agents run on your own AI plan — a Claude subscription, ChatGPT/Codex, or a model you connect through OpenCode — paid on the provider’s site. The free edition of harnsy runs 2 projects in one organization; harnsy Max, the paid edition, adds more projects and organizations. More agents at once use the plan’s limits faster; the Usage page shows how much is used. Their sites show the prices. harnsy doesn’t sell model access.

All questions →
