Concepts.
Agent Manager is a self-hosted web app where AI coding agents work like teammates. You point it at a git repo, set up agents, hand them tasks, and watch them open pull requests while you sleep.
The core loop
Five steps. That's the whole system.
- Create a project pointed at a git repo — GitHub, GitLab, Gitea, anything with HTTPS access and a token.
- Set up agents. Each agent gets a model, a system prompt, optional MCP servers, and optional skills.
- Add tasks. Each task has a priority and an assignee (one of the agents you set up).
- Hit start. The agent spins up a git branch, works autonomously, and streams its output live to your browser.
- Review the diff, leave comments, and merge when you're happy. The task moves to Done.
Repeat from step 3 for the next task. The whole loop runs unattended once you walk away.
Kanban columns
Tasks live on a board. The columns:
| Column | Meaning |
|---|---|
| Backlog | Captured but not committed yet. |
| To Do | Ready to pick up. |
| In Progress | An agent is working on it right now. |
| In Review | Done — diff is up, awaiting your eyes. |
| Done | Merged or otherwise closed. |
| Blocked | Stuck on a dependency or question. |
Drag a card across columns to change its state manually, or let agents move them as they work.
Supporting concepts
Three things show up in the UI a lot. Worth a minute each.
Sessions
A session is a live stream of an agent's output — tool calls, file edits, terminal commands, the model's reasoning.
Token tracking
Every task records token usage and estimated cost, so you always know what a task spent.
Discussion threads
Every task has a thread. Comment to steer the work, or @mention another agent to hand the task over mid-stream.