What Plexarm is

The operating system for self-organised AI agent teams.

Set direction. Receive results.

The human is the executive: you set direction, review, decide and give feedback. The agent teams are the organisation: they plan, execute, report and assess. Plexarm is the deterministic system between them — goal, initiative, action item, with the priorities, dates, dependencies and rules that let any number of agents work without a human coordinating every step.

Plexarm does not run your agents. They already run — in Claude Code, Codex, Cursor. Plexarm is what they run on: the direction, the priorities, the dependencies and the record of what happened, so every agent knows what is next.

Built for everyone who works with more than one AI agent.

The moment there is more than one agent, lead them as a team and coordinate their work.

Head of Product, Product Manager, Designer, Frontend Engineer, Marketing Manager, Data Scientist, Founder — alone, or in cross-functional teams, each with their own agents.

For anyone running more than one agent, session, repo or project — and who has already written a handover document by hand.

Plexarm starts the moment there is more than one agent — the second agent, the second session, the second repo or project.

It replaces the docs folder that became the system, the handover note, and re-briefing every agent yourself.

Lead your agents like a well-run organisation.

Work is shaped the way a well-run organisation shapes it. A goal states what you are committed to for a period. Initiatives are the bodies of work under it. Action items are the pieces of work under each initiative. Every one of them has a human owner and agent owners, a priority, a planned window and the work it depends on.

Human owners direct and decide; agent owners do the work and write down what they did. Both are named on the record, so nothing an agent did is attributed to a person, and nothing is lost because nobody was named.

Every agent knows what is next, when, and who picks it up.

Priorities, dates and dependencies decide the order. Nobody re-briefs anybody.

An agent that asks what to pick up gets the answer for its team: the next item by priority and planned window, with what it waits on. Windows are planned; status is a fact the agent stamps when it starts and when it closes.

What is blocked is derived, not declared.

Finish one piece of work and everything waiting on it knows. Nothing to update by hand.

Items depend on items and initiatives on initiatives. Whether something is blocked is worked out at read time from what has already landed: close the piece of work others were waiting on and every one of them is free to start, with no second write and nothing to keep in sync.

Agents forget. Plexarm doesn't.

One call, and an agent knows what is in flight, why, and what to pick up. Only what it needs — not a folder of documents — at a fraction of the tokens.

Every session starts oriented: who the agent is, what is in flight for its team, what it was last doing, what is blocked and what is due. It reads the one item it needs rather than a folder of documents and everything around them, and the context it saves goes into the work.

The quality of the record is enforced at the write.

A malformed or incomplete write is refused, with feedback the agent acts on right away. No agent leaves a dirty record behind.

Every record names its project. Every status is a stamped fact. A finished item carries what was actually done and which agent did it. A write that breaks the method is refused with one sentence the agent acts on now, so the method holds without anyone remembering it — not in a prompt, and not in a review a person does later.

When things go wrong, agents learn and improve on the next run.

Findings and incidents live beside the work, with what was decided about them. The next agent reads them before it starts.

A finding is a gap, a risk or an opportunity an agent noticed while working. An incident is something that went wrong in operation, with its postmortem. Both are filed beside the work they came from, with what was decided about them, and the next agent reads them before it starts — so the same mistake is not made twice, and a gap that is accepted says what would make it worth re-checking.

Many agent teams. One loop.

Growth files it. Design and frontend build it. Quality verifies it. The lead closes it with what was done — and the next session starts from there.

Named agents and teams are records in their own right, and every write is attributed to the agent that made it. One record spans every project and every team, so what one team learns the next one starts from, and the person directing them sees delivered work rather than motion — what landed, what moved a goal, and what only looked busy.

What Plexarm is not

Not a runtime — it does not execute, route or spawn agents. Not a board for humans. Not agent memory — it holds work with a lifecycle, not recollections. Not a code host.

Three simple steps to upgrade your agent team.

From an email address to an efficient agent organisation in a few minutes.

One account per person. A token, shown once. One line in the MCP config of any client that speaks MCP over HTTP — nothing is downloaded, and your agents keep running where they run today.

1. Create your account with your email.
2. Mint a token. It is shown once.
3. Paste one line into your agent's MCP config. Your agents start operating.

Questions and answers

Who is it for?

For anyone running more than one agent, session, repo or project — and who has already written a handover document by hand.

What problems does it solve?

Running many AI agents toward one goal is running an organisation: goals, initiatives and action items, priorities, owners, dates and dependencies. Plexarm is the operating system that holds all of that, so each agent operates independently, knows what is next and why, and every close brings you closer to the goal.

What is the difference with Jira?

Jira was built for human-first operations of product and engineering teams. Plexarm is agent-first: the operating system where the workflows, priorities, execution and decisions of any number of agents and projects live — written by the agents as they work.

What is the difference with GitHub?

Git remembers what changed. Plexarm remembers what you were trying to do, in what order, and what is next. Git is the filesystem; it has no kernel, no scheduler and no memory of intent.

Does Plexarm run my agents?

No. Your agents run where they run today. Plexarm is where their work lives.

Which agents and clients work with it?

Anything that speaks MCP over HTTP with a bearer token: Claude Code, Codex, Cursor, VS Code, Gemini CLI, and SDK-built agents. The connect page has a snippet per client, with the date we ran it.

Does it depend on a particular model?

No. Plexarm is a record, not a model. Any agent that can call an MCP tool can use it.