Running several coding agents at once turns into babysitting: one tab per agent, each waiting on you at a different moment, and no single place that says what needs a decision. Collie is the herdr plugin that takes that coordination off your hands. You pick a workflow, Collie opens the tabs, runs the agents, reviews their work, and brings you back only for the decisions that are yours. One board shows every task in the session: what needs you, what is working, what finished.
Run one command. It's safe to run again, and it ends by telling you whether you're ready or what's still missing and how to fix it:
git clone https://github.com/cego/collie.git ~/.collie && ~/.collie/setup.shBefore you run it, you need:
- herdr installed and started once. See the herdr install guide.
- A
glab auth logintogitlab.cego.dk, so the workflows can open merge requests there. - Claude Code, logged in. The bundled workflows run their agents on it.
Nothing from GitHub: the repository and its releases are public. collie onboard installs
herdr, Claude Code and Collie at a release instead, one reported step at a time; see
Onboarding a Machine.
The install links the plugin, puts collie on your PATH, adds four keybindings, installs
the required skills, and runs collie doctor. Later, collie upgrade brings all of it up
to date. Helle and a Linear MCP are optional; doctor tells you when a workflow you'd run
needs one. See Install and
Optional integrations for the detail.
| Key | What it does |
|---|---|
prefix+f |
Run a workflow |
prefix+u |
Resume a run that stopped |
prefix+shift+f |
Fork a persona into your layer |
prefix+shift+c |
Open this session's Control Plane |
prefix is ctrl+b by default. Each herdr session has one Control Plane in a workspace
of Collie's own: a board with one card per task, grouped by what needs you, what is
working, and what finished, with a chat pane beside it where you can ask about the work
or redirect it. prefix+shift+c reaches it from any pane and opens it when the session
has none yet.
| Workflow | What it does |
|---|---|
plan |
Interviews you about the goal, writes SPEC.md and tickets into the run directory, then offers: implement now, second opinion, offload to Linear, refine |
implement |
Builds from a plan, a Linear issue, or a description on a branch of its own, has the change reviewed, fixes the findings, then opens the merge request |
review |
Reviews a merge request, a branch diff, or the working tree, writes one review, and offers to post it |
architecture |
Runs the architect over the project, reports into the run directory, then offers: implement now or stop |
renovate |
Merges the month's Renovate merge requests on one repository, tags a release, and checks the repository off the team's Renovate issue |
The bundled workflows are a shared starting point, not a restriction. collie workflow fork <id> writes a module of your own that imports what it keeps and replaces what it
changes, and a workflow you write yourself is found where you saved it. A persona forks into
your own layer with prefix+shift+f, changing only the keys you disagree with.
Every action is also a command, so an agent can drive Collie:
collie run start review --input target=worktree --input plan= --input proves=- Using Collie: install, keybindings, the Control Plane, hand-offs, your defaults, troubleshooting.
- Workflows: what each workflow is for, what it needs, how they chain.
- Authoring: where a workflow is saved, forking, personas and their frontmatter.
- CLI: commands,
--jsonenvelopes, exit statuses, idempotent retries. - Internals: architecture, the host, build and release, running the tests.
CONTEXT.md: the vocabulary.- ADRs: the decisions and why.
AGENTS.md: where an agent working on this repo starts.- Brand guide and brand showcase:
logos, icons, color, typography, and usage. Open the showcase in a browser, or publish
the
assets/directory with any static host; it has no build step.
Where Collie is going: PRODUCT.md.

