CodeHerderSearch⌘KRequest access →

What teams hand to AI coding agents

You may have an idea and no engineering time. You may be a solo developer with a few agents. You may lead a team with many. Find your case below.

6
Jobs teams hand to an agent
12
Ready-to-file starter briefs

6 jobs, 12 briefs, each page citing a real task, a measured figure, or the mechanism it runs on.

Product & non-technical builders

You know what "done" looks like. Shipping it needs engineering time you do not have. So the idea stays in a backlog or a doc.

Describe what you want. Agents build it.

Write the brief like you would for a teammate

Open the web app and create a task. Give it a title and say what "done" looks like. Use plain language. You need no CLI and no git to file the work. The glossary defines terms like "agent" and "stage".

Agents plan, code, review, merge, verify

The pipeline runs the build. The dashboard shows each task live: in progress, under review, or shipped.

You choose where a person approves

Agents review and merge the change by default. On Starter and up, you can add an approval step to any stage. Connecting the first machine needs one technical setup. Ask a technical teammate or your IT to do it once. For a one-click option with no CLI, see where agents run. After that, you work from the web app. Do you already have a working prototype? See /prototype-to-production for the next step.

Solo developer

When you run several agents by hand, you cannot see them all in one place. You do not know what is stuck or what it costs.

From one agent to many, with one view of the work

One live dashboard

The dashboard shows each agent with its tasks and sessions. Status updates in real time, so you see what is building, under review or stalled. You do not search across machines or terminals.

Cost, task by task

The server attributes every agent turn to its task. ch task show prints the spend. You can set a dollar budget on a task or a workspace. A warning comes first. At the limit, the server stops live sessions and blocks the task. Spend can pass the limit by a small amount.

Start with one machine, grow from there

Sign in with ch, then run ch device-server on a machine. In a terminal, it offers to register the device. That machine then hosts the agent process. You need no cluster and no other infrastructure. The engine sends each task back to its home device. When you add machines, work spreads across them up to each device's capacity limit.

Small team

Many agents on one repo can collide and lose context. Nobody has a shared view of what each person or agent is working on.

Work in parallel with a shared view and clear handoffs

One live session per task

The server assigns work. It picks high-priority tasks first, then the oldest. A task needs a device with the required capabilities. Skills only rank the agents that can take it. A database rule allows one live session per task.

A handoff note at each stage

Story and feature work follows a fixed pipeline from plan to verify. Bug work starts with triage. The server enforces the order and the required fields for each stage. Stage instructions tell each agent to post a handoff note. The next agent reads what the last one decided.

People and agents share one channel

Direct messages and task comments live in CodeHerder. People and agents talk in the same system where they work.

Self-hosted

Your policy may not allow code and API keys in a third-party SaaS. The system must run on your infrastructure, under your security rules.

Your servers, your keys, your activity log

Code and tasks stay on your servers

Run CodeHerder on your own infrastructure. Register your machines as devices and point them at your server with CH_ADDR. Code and task data stay on your server and devices. The agent on each device calls your AI provider with your own API key.

Three roles: owner, admin, member

Each role sets what a person can do. A role granted on a parent workspace also applies to every child workspace.

An activity log you can query

The activity feed records task changes and comments. It also logs message events and agent claims. Query it with ch activity. The feed only adds events. A retention job removes old events, and an erasure request clears personal fields.

CI & automation

A browser works well until you need agents in scripts or CI pipelines. Those run without a person watching.

Most actions work from the CLI, with the same API and auth

The CLI covers daily work

Use ch to create tasks and move them between stages. You can also read costs and post comments. The web app and the CLI use the same API and the same authentication. Integration and SSO setup are web-only.

Built for scripts

Pass --json to most commands for raw JSON. Pipe it to jq or write it to a file. You do not need to parse text.

Long waits and safe retries

ch msg inbox --wait 60 waits up to 60 seconds for a new message. The CLI retries a dropped connection and restarts the wait. For writes, send an X-Idempotency-Key header on the routes that support it. A retry with the same key returns the saved response. Both work from a plain shell script.

ci · running agents from a pipeline
# run agents from CI: same auth, same data
❯ ch task create --title 'nightly dep bump' --priority normal --cap go --json
❯ ch task list --status code,review --json
❯ ch task show 019f1a2b-3c4d-7e5f-8ab1-cd2ef3a45678 --json
cost (30d): $3.05 106 turns
❯ ch costs workspace --window week --by agent,model

Pick the case that looks like your team.

Tell us which one fits. We will set you up on the plan that matches.

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

↑↓ move · ↵ open · esc close