Mainplane: Open-source agent runs multiple PCs, skips context compaction

What this is about

Mainplane is an open-source "agent harness": the software shell around a language model that turns a plain text generator into a worker that calls tools, edits files and completes tasks. It was developed by Alexander Yue and, according to a Japanese overview by the author npaka, released on October 26, 2026. It's implemented in Go and licensed under the MIT License, so you can use and modify it freely.

Here's how it differs from typical AI coding agents: those usually run on the machine you're sitting at. Mainplane is designed so that an always-on server controls multiple PCs. It also takes its own approach to context management, which often becomes a problem on long tasks.

Server, worker and multiple PCs

Mainplane separates server from worker. One example from the source: a Linux machine at home runs permanently as the server, while your Mac or Windows development machine connects as a worker. Both can be linked over the internet or a private network, so you can hand tasks to remote PCs too.

There is also a shared file system that several machines can use at the same time. If you want to host it, though, you need a suitable Linux machine. Multiple PCs aren't required: server and worker can run on the same machine.

Agents and tools

The agents work by having the language model (LLM) call tools. Mainplane deliberately keeps these tools simple and leaves it to the model to decide how to use them. According to the source, that covers not just code changes but also researching information in a browser and operating programs with a graphical interface. You can manage multiple sessions and choose the model and target PC for each task.

On the model side, Mainplane isn't tied to one provider. According to the official README, it works with API keys as well as Claude Code and ChatGPT subscriptions, within their usual usage limits. Through so-called connectors, external services can send messages to the agent and kick off tasks, so you don't have to reach it only through the terminal.

Context retrieval instead of compaction

Agents such as Claude Code or Codex often use compaction in long conversations: the model summarizes the history so far, and work continues from that summary. That costs time and tokens (the billing unit for text), and details can get lost.

Mainplane instead stores the complete original history and pulls out only the information needed at any moment ("context retrieval"). The developer compared both approaches using Claude Opus 5.5 on a session of about 400,000 tokens. Mainplane came out faster and cheaper. The source doesn't give concrete figures, and it stresses that these are measurements under particular conditions that won't necessarily hold for every task. They're also vendor-reported numbers, not independent tests.

Security and privacy

Because one agent can operate several machines, access permissions matter a great deal. If you allow shell commands or file operations, unwanted changes or data leaks are possible. According to the README, user data isn't sent to the developer by default; data sharing is optional. However, the default configuration uses an external service to find connection targets, and Cloudflare Tunnel for communication. On top of that, data flows to your LLM provider, whose privacy rules apply separately.

Context retrieval has limits too: it doesn't guarantee that all the necessary information is found. Where knowledge is scattered across several earlier sessions in particular, a weak search can miss important details. Since the project is new, the source says its stability and long-term track record aren't yet proven.

Background: why the harness matters

A Forbes piece by Rohan Pinto (1Kosmos) summarizes a study that took apart 11 coding agents. All of them consisted of the same seven parts: loop, model layer, tools, memory, safety controls, orchestration and extensions. The thesis: agent = model + harness, and a better model raises the ceiling while a better harness raises the floor. Memory is exactly the area Mainplane targets. Pinto recommends, for example, deliberately deciding what the agent keeps and what it forgets, since long histories slow things down and cost money.

One clarification: the company name Harness, which appears in two other sources (a code repository with AI review, "Agent DLC"), refers to a software delivery platform and has nothing to do with Mainplane.

Our take

Mainplane is interesting for developers who want to keep agents running around the clock on a home server and tie in several computers, without being locked into one model provider. The approach of querying histories instead of summarizing them addresses a real problem in long agent sessions. How robust it is in practice remains an open question: the benchmark comes from the developer, the source gives no numbers, and the project is only a few days old. If you try it, keep permissions tight, work in containers or on test machines, and remember that connections go through external services by default. There are no model weights here. This is a tool, not a model release.