CodeHerderSearch⌘KRequest access →

What AI coding agents cost, and how you control them

You do not need to read a diff to run this. You need to know the cost and the quality bar. You also need to know who is accountable and where it stops.

$7.48
Mean cost per shipped story
84.8 min
Median time from brief to merge
16,343
Agent task branches merged into main

Cost and lead time measured across 280 shipped stories, the week of 9–15 November 2026. Task count as of 20 November 2026.

Five questions to answer before you approve

Your board or your manager will ask these. Each answer links to its source.

What does it cost?

The mean cost of a shipped story, rework included, is $7.48. CodeHerder estimates the cost of each task from token use and model list prices. A review round or a failed check shows in that number. The Costs page splits spend by agent or by model. You can also split it by stage.

Four spend caps work together. Only spend billed to an API key counts.

  • A daily cap. One account value applies to every agent.
  • A default cap on each task, set once for the account.
  • A budget on one task. It overrides that default.
  • A budget on a whole workspace. It counts all spend since the workspace began.

CodeHerder warns you at 80% of a cap. At 100%, a task or workspace cap stops live sessions and blocks the task. The daily cap only refuses new sessions. The account-wide caps and the workspace budget need the Pro plan.

Is the work any good?

CodeHerder has six server-enforced gates. Each workflow turns on its own set:

  • Required fields
  • Required verifications
  • Checklist gates
  • Approvals with separation of duties
  • Dependency and close gates
  • Change-size limit

The default story workflow uses required fields and a checklist. It also needs a passing verify check. Approvals and size limits are opt-in on a custom workflow.

CodeHerder runs on its own gates. This codebase carries 45,032 go test functions in the suite, built the way your work would be.

Who is accountable?

Agents review and merge the work by default. No person approves it. To require one, use a git-host approval. The merge stage tells the agent to block the task when that approval is missing. This holds only if your host denies the merge identity approve and bypass rights.

You can also add an approval gate to the entry stage of a custom workflow (Starter plan and up). A member you choose then approves each task before an agent starts. A checklist gate blocks a move past the proof stage of any unticked acceptance item. It applies to workflows with a checklist. The story workflow has one. The activity feed records each status change and each comment. It records each blocker too.

What is the risk?

Each task gets its own git worktree and branch. Overlapping edits raise an alert, and the merge agent resolves conflicts. Agents run on devices you own, or in your own AWS account. We do not run them. Billing goes to your own AI account, through an API key or a subscription login. A self-hosted deployment keeps coordination inside your own perimeter.

Does everyone deliver work the same way?

Without a shared process, each engineer runs agents their own way. Checks and hand-offs depend on who runs the agent, and progress is spread across chats and terminals.

In CodeHerder the team defines its workflows once. Each stage's prompt lives in the shared stage library and is versioned. Skills and the workspace wiki reach every session. The server enforces each workflow's gates, whoever filed the task. Set these in a group and every workspace under it inherits them. Custom workflows and stage prompts need the Starter plan.

What you watch each week

Four places show how your agents are doing. None of them need you to read code.

  • Dashboard spend

    The Costs page splits spend by agent or by model. You can also split it by stage. The task list shows the cost of each task.

    Understanding costs →
  • The stalled count

    Every task that stopped moving, at any stage. Click the tile to open that list.

    Tracking work →
  • The Review debt panel

    How long work waits for review. How long review takes. How often work bounces back for rework.

    Review debt →
  • The Activity feed

    Each status change and comment shows here, newest first. Blockers show here too. Filter by one task or the whole workspace.

    Activity feeds →

Who gets unblocked

Developers

Hand off routine work. Your team spends its time on work only a person can do.

See the use cases →

Product people

Brief a task in plain words from the web app. Your engineers keep the review rules. After a one-time device setup by a technical person, you need no CLI and no git.

For product teams →For developers →

You

See every agent and every task in one dashboard. Spend shows there too, so you can skip the status meeting.

Start with a two-week pilot

The plan runs from connecting one device to pricing month two from what you saw. The two weeks are a guide. Your speed depends on your repository.

  1. Day 1

    Stand it up

    Owner: Technical owner

  2. Day 2

    Hand over one small task

    Owner: The person with the idea

  3. Days 3 to 5

    Read the gates before the diff

    Owner: The approver

  4. Week 2

    Run three tasks at once

    Owner: Technical owner

  5. End of week 2

    Decide on your own numbers

    Owner: The person with the idea

Read the full plan on /rollout. Copy a task from /starter-briefs to run phase two today. To compare the plan with hiring a person, read freelancer or agents.

Where it stops

These limits apply today.

  • A host besides GitHub or GitLab does not work yet. The merge stage signs in with your own gh or glab.
  • Keep back any change you cannot check. An example is a change with no test and no acceptance criteria.
  • By default, no person approves the work. To require one, add an approval gate or a git-host approval.
  • Access to the hosted service is invite-only right now.
  • Gates catch most mistakes before they ship. They do not catch every one. See what happens when one gets through.

Questions leaders ask before they sign off

How do I forecast spend?

Price your own volume on the calculator, using our measured cost per story. CodeHerder estimates cost from token use and model prices, so treat each figure as an estimate. Set a spend cap on a task or a workspace. CodeHerder warns you at 80% of the cap. The workspace budget needs the Pro plan.

Can an agent merge without a person?

Yes, by default. In the story workflow, the review stage still checks the code before merge, but no person approves it. To require a person, use a git-host approval. The merge stage tells the agent to block the task when approval is missing. That holds only if your host denies the merge identity approve and bypass rights. You can also add an approval gate to a stage in a custom workflow. A checklist gate blocks a move past the proof stage of any unticked acceptance item.

Where does our code run?

On a device you own, or in your own AWS account. We do not run it. Each task gets its own git worktree and branch. A self-hosted deployment keeps coordination inside your own perimeter.

What do I need from my team to start?

You need one device and one repository. You also need a technical teammate for the one-time setup on day one. Filing and approving tasks after that needs neither.

How do I know it's working after two weeks?

Read what each task cost and what it shipped. Then price month two from your own numbers.

Bring one team, one repository, two weeks.

Connect a device and file a first task. Then read the cost and the diff yourself.

Try "pricing", "connect a device", or "who reviews the code"

↑↓ move · ↵ open · esc close