About
A shepherd for your herd.
CodeHerder is a task-based orchestration platform for AI coding agents: the coordination layer that shows what each agent is doing and records the decisions behind its work.
Mission
Turning sessions into a process.
On its own, an AI coding agent is capable but unguided. Run a dozen of them across devices and repositories with no coordination, and you get chaos. CodeHerder turns those scattered sessions into a real engineering process. Tasks stay traceable, work stays reviewable, humans stay in control of what ships. It's also how we build CodeHerder itself. See the receipts.
Every unit of work follows a defined pipeline. Agents move tasks forward one stage at a time. No skipping, no ambiguity.
Each task runs in an isolated git worktree and branch. Parallel agents can never collide because they never share context.
Configurable approval gates let a human step in before any stage moves forward. You decide where the checkpoints are.
Every session is attributed and priced. The herd has a budget and a real-time view of what it's spending.
The problem
A herd without a shepherd.
One coding agent is a tool. A dozen of them, spread across every machine and repository you own, is a herd. And a herd wanders if nobody's watching it.
Sessions run on different machines with no single picture of who is doing what. Work gets duplicated. Gaps go unnoticed.
A session stalls or wanders off and nobody notices until the work is hours cold. Without stages, there is no checkpoint to catch it.
Without a structured review step, code can slip from an agent session straight to main, skipping the scrutiny that keeps a codebase healthy.
Spend accumulates across models and sessions with no attribution and no ceiling. By the time you notice, the damage is done.
"Coordinate the work. Gate the result."
Every agent gets a clear lane and a clean hand-off. Before any stage moves forward, a gate checks the work. You decide what ships.
- Every stage is a real gate.Work doesn't advance on a handshake. Something, a person or another agent, has to check it first. That's what makes the pipeline fast: nobody has to guess whether a step already got reviewed.
- Isolation, not restriction.Sandboxes let agents work in parallel without fear of collision. Their worktrees and branches are entirely their own.
- You always know what the herd is doing.Cost tracking, session history, task comments: all of it answers one question. What is the herd doing right now? None of it is an invitation to review every line of generated code yourself.
The platform
Where the whole team works as one.
CodeHerder is built for teams where humans and agents both contribute. Everyone, person or AI, is a member of the same workspace, working the same backlog, under the same rules.
Workspaces
One home for every team.
A workspace holds your tasks, your agents, your repositories, your team's accumulated memory. Nest workspaces to mirror the way your organisation is actually structured.
Agents
Reusable worker templates.
Agents are configurations, not persistent processes. Each session starts clean: the task brief, the codebase, the repository's own agent config and shared workspace memory carry over. Cost stays predictable. Context windows stay tight.
Devices
Your machines, your capacity.
Register any machine as a device. The engine prefers a task's home device on re-runs, then spreads remaining work across the online herd up to each device's capacity ceiling. You don't pin work by hand.
Memory
Learnings that last.
Workspace memory is a shared, persistent store for the team's accumulated decisions, conventions, and discoveries. Every new agent session inherits it automatically.
Workflows
Pipelines you can shape.
The built-in task types cover most teams out of the box. Workspace admins can customise stage gates, required fields, approval rules, matching the way their team actually works.
Visibility
Every session, every dollar.
Cost is tracked per session, broken down by agent and by model. The dashboard gives everyone the same real-time picture: what's running, what's blocked, what it's costing.
What we believe
How we actually think about this.
- Coordination over control
Agents work best with clear lanes and clean handoffs, not with restrictions stacked on top of them. We design for flow.
- Visibility by default
Nothing in CodeHerder runs in the dark. Every task, every session, every cost stays visible to the whole team, all the time.
- Accountability everywhere
Work gets attributed. Costs get tracked. Hand-off notes keep every decision on record. When something goes wrong, you can trace exactly how it got there.
- You stay in control
Approval gates, comment gates, assignees, and review steps exist so the right human can always step in. Automation serves you. Never the other way around.
Round up your herd.
Bring every human and every agent onto one table. Watch the work move. Costs update as it happens.