Image: Herdr Projects
Why it mattersA developer who already runs Claude Code, Codex or pi can now hand many parallel tasks to agents and stay in a yes-or-no loop on the ones that need an answer or a merge.
Briefing a coding agent, waiting for it, reading the diff, approving the merge, then starting over with the next task takes most of the saving the agent was meant to give. Herdr Projects, a plugin for the Herdr agent terminal, puts the ones you delegate on one board and lets you type what needs doing instead of briefing each agent by hand.
The project passed 740 stars in 30 days as of 2026-10-16 and shipped version 1.0.0 on 2026-10-15, by Elias Stravik. It is MIT licensed and runs on macOS and Linux with Herdr 0.9.3 or newer. The author's own description: "one list of tasks, yours and the ones you delegate. You type what needs doing and the board's own agent, the Coordinator, turns it into tasks."
What the board does
Each delegated task runs as an ordinary Herdr agent in its own git worktree on a branch named work/<n>-<title>, so the user's own checkout is never touched. The board groups rows into five sections in the order they need a person: needs you, ready to approve, working, to do and done. The author says the header counts them in a short line, for example 3 need you · 1 ready · 2 working.
A finished task walks a fixed path before anything lands on main. The agent commits its work, the board merges main into the branch, the repository's own check runs, a draft pull request opens, and a second agent reviews the result. A failed check or review goes back to the task's agent. Work that passes waits for the user to press Ctrl+L before it merges.
Agents and permissions
The board starts each worker on an agent CLI the user already runs: Claude Code, Codex, pi or another agent Herdr supports. It passes the task and the board's rules, and the author's README says it runs the agents with full access by default (claude --dangerously-skip-permissions, codex --dangerously-bypass-approvals-and-sandbox) so a task does not wait on a prompt no one will see. The documented switch is worker.access ask, which starts the agents without those flags.
The board also keeps its files on the user's machine, in ~/.herdr-projects/ by default, with its config in settings/config.json, a folder per task under data/tasks/<repo>/<n>/, and the Coordinator's chat history under data/coordinator/. It reaches other machines over the user's own SSH. The author's README says review links go to the user's tailnet or 127.0.0.1.
What a reader does with it
A developer paying for Claude Code or Codex and running several tasks at once gets one place to see which agent is waiting on an answer, which has work ready to approve, and which has drifted. The author's own documented limits are honest about the trade: a review is a second agent's judgement and never a proof, an automation runs with no one watching, and a repository's own config cannot loosen a guard the board owns.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.