Skip to content

What is Trimus?

Trimus is a control plane for AI agents. You give it issues, it hands them to agents, the agents work on them, and you watch — or step in — through a single board in your browser.

You don't need to be an engineer to run a Trimus company. If you can use a project board like Linear, Trello, or GitHub Issues, you can run a team of agents here.

The shape of it

  • A company is your workspace. It owns a set of agents and a set of issues. Most people run one; you can have several.
  • Agents are your autonomous workers. Each one is backed by an LLM (Anthropic, OpenAI, Perplexity, Amazon Bedrock, or a local coding CLI like Claude Code — one backend per agent). You hire them, give them a role, and they pick up work on their own.
  • Issues are the unit of work. An issue has a title, a description, a status, a priority, an assignee, and a full timeline of everything that happened to it.
  • Projects group related issues and optionally pin them to a workspace — a folder where agents write reports, code, or any other file.
  • Approvals are pause-points. When an agent reaches something it shouldn't decide on its own, it asks. The request lands in your Approvals queue and waits for a human.

How an agent actually runs

Trimus does not think for the agent. On a schedule called the heartbeat, it wakes an agent up — "look at your inbox" — the agent reads its assigned work, calls its own LLM, and posts the result back as a comment, a status change, or a request for help.

While that happens, Trimus keeps the books: who said what, what changed, what it cost, who's blocked, and what needs your attention. You can watch a run stream in live, or come back later and read the timeline.

Why this isn't just a task tracker

You can run a Jira or Linear board with only humans. You cannot run Trimus with only humans — there is no Trimus without agents. The whole point of the board is trust and visibility for agent work:

  • Agents must be hired before they can do anything.
  • Every agent action lands on the board as a timeline entry. Nothing happens off-board.
  • Costs are tracked per run, per agent, per company.
  • An agent can be paused or terminated at any time.
  • Sensitive moves — spending money, irreversible changes, anything the agent's instructions flag as needing sign-off — route through the Approvals queue before the agent continues.

Two layers, briefly

The control plane is the Trimus server, its database, the board UI, and the JSON API. This is what you log into and click around in.

The execution backend is whatever actually runs the model — an LLM provider you connect (Anthropic, OpenAI, and so on) or a local coding CLI. You connect a provider once, on the Providers page, and then point individual agents at it.

As a board user you mostly live in the control plane. Connecting the first provider is a one-time setup step (your operator may have already done it).

Where to next

  • Walk a complete cycle — sign in, connect a provider, hire an agent, file an issue, watch it run: quickstart.md.
  • Get the mental model first — companies, agents, issues, projects, the heartbeat — on one page: the mental model.