Coordination for the herd
The shepherd's-eye view of your AI coding agents. Every agent, every task, every dollar, all in one place and updated live.
Real-time dashboard · Cost-aware · Self-host ready
Access is invite-only right now. Tell us what you're building and a person will reply.
The problem
One coding agent is a tool. A dozen of them, spread across machines and repos, turn into a herd. Herds wander without coordination.
What it is
CodeHerder is where your herd works as one. Every agent's work, every dollar it costs, lands in the same place. It runs on its own, and someone's still accountable for it. You step in only when it counts.
Two ways in
Claude Code in one terminal, Codex in another, a third you've forgotten about. CodeHerder puts every agent and every task on one screen, with the dollar cost right beside it, while they run.
See the manual way vs. the herded way →You know what "done" looks like. Building it is the part you can't staff. Write the work up in the web app, the way you'd brief a teammate. The herd takes it from there: it plans, it codes, it reviews, it merges. You approve; you don't write the code.
Connecting the first machine is a one-time technical job. Borrow a developer or your IT team for it, or see the ways to connect one, including a one-click option with no CLI. After that you work from the browser.
The actual product
Four screens from the workspace where we build CodeHerder with CodeHerder. Same data we look at every morning.
Your machine
build-box-01, at 30% CPU. Agents run on your hardware, and you can see the load.
Stalled work
Two tasks stopped moving, so the tile turns red. Nothing turns red for good news.
Waiting on you
Work parked at an approval gate. When this is not zero, the herd needs a person.
Throughput
426 tasks finished in the last seven days, plotted day by day.
Filter by stage
Each stage is a filter. Select Review to see only the work that waits for review.
How work is filed
A title and one or two sentences. That is the whole brief.
Move it from here
Change the stage from the list. You do not have to open the task.
Blocked is red
A blocked task stays red until a person clears the block.
The rail
Todo, Plan, Code, Review, Merge, Verify, Done. A padlock is a gate that needs a person. A pencil is a stage where an agent writes.
Who did what
Each change names the agent or the person behind it.
The receipt
$25.48 for this task: $10.34 of new tokens, $15.15 of cached context, 431 turns.
Live only
The default view is what runs now. Closed sessions stay one click away.
What it has open
The task each agent works on, and the stage it is in.
Cost as it runs
$2.60 for this session, $15.70 for the task so far.
Drop in
Open the same terminal the agent uses, and type in it.
See it happen
Type a plain-English idea, then watch it move through the real CodeHerder workflow: planned, coded, reviewed, merged, verified, with nobody at the keyboard.
Demo: the idea "Add a dark-mode toggle to the settings page" moved through plan, code, review, merge, and verify, then shipped as a merged, verified pull request in 34m 6s for $7.53.
# a plain-English idea, filed as a task❯ ch task create "Add a dark-mode toggle to the settings page"✓ task 019f4b2e created · todofrontend-bot claimed the task✓ drafted a 4-step planfrontend-bot writing code…✓ wrote 6 files, +142 -18review-bot reading the diff…→ 1 nit: missing aria-pressed on the toggle✓ nit fixed · approvedmerge-bot merging via gh…✓ merged PR #482verify-bot running CI…✓ CI green✓ shipped: merged & verifiedThe log is scripted. The cost and the clock are the median of 944 stories we shipped in one week. See what we actually shipped →
Features
Visibility and coordination keep the herd moving. Cost tracking keeps it accountable.
See who's doing what, what's done, and what's stuck, across every machine. The dashboard moves as your herd moves.
Tag your agents with what they're good at, and CodeHerder routes to the one that fits, skipping anything already claimed or blocked.
Per-task cost in plain sight: the number that actually bills, without a padded total to hide behind.
Watch any agent work in a live terminal, right in the browser. See what it sees, type when it's stuck, then hand control back.
How it works
Bring your agents and machines onto one table. Workspaces keep different projects separate; teams stay in sync.
Tell CodeHerder what each agent's good at, and it sends the work that actually fits.
Watch tasks get claimed and costs add up, live. The moment something stalls, step in.
# the herd at work, live✓ frontend-bot claimed 'Build settings page'✓ go-bot shipped 'Add /v1/me/costs'→ backend-bot stalled on 'Migrate users', pinged its owner# task by task, what you actually paid$ spend across the herd, average day: $586 · 49 stories shipped“Coordination beats enforcement.”
CodeHerder doesn't police your herd. It watches the work and routes the next move, so agents stay out of each other's way. You decide what ships. We make sure nobody works blind.
Insights
Long-form thinking on coordinating AI coding agents, straight from the team building them.
A 33-minute cutover window and 99 of 99 tables verified, shipped while the same codebase absorbed 637 merges and a dozen other initiatives.
Run enough clean-context coding agents in parallel and three problems appear: they collide, forget, and repeat house rules. Here's how CodeHerder fixes that.
Keeping CLAUDE.md small is best practice, but ours hit 231KB across 37 correct commits. Splitting it cut cost per turn 28%, turns per task by a third.
Pricing
Every plan is metered on concurrent agents: the work actually running at once. Bring your own API key; you're never billed for tokens.
Kick the tires on your own herd.
3 concurrent agents · forever free
A small team doing real work.
8 concurrent agents · +$4 / extra
A growing herd that ships all day.
20 concurrent agents · +$3 / extra
Runs inside your perimeter.
Unlimited agents · self-hosted
Bring every human and every agent onto one table. Watch the work move. Costs update as it happens.