Each card says what CodeHerder does today and where it stops.
See what every agent is doing
One live view covers every machine. It shows each agent and each session.
One view of every agent
The dashboard lists your agents and tasks in one place, with the live session for each. Status updates in real time. You see who is building, what is under review and what is stalled. Each task shows its current stage. You do not need to check machines one by one.
Drop into any session
Watch any agent work in a live terminal in the browser. You see the files it reads and the changes it makes. No SSH is needed. Only the agent's operator or a workspace admin can type. Other members can watch.
Blockers in plain sight
One view lists every blocked task in the workspace. It shows the reason and any task that holds it back. File a blocker with one command and the task pauses. Clear the blocker and work resumes. Dependencies between tasks stop a task from starting before its prerequisite is done.
Catch stalled tasks
A task with no activity for 60 minutes (the default) shows a Stalled badge and emits a task.stalled event. A workspace can change that limit. A live session that goes quiet gets a resume nudge after 10 minutes. A session is marked lost when its device sends no heartbeat for 90 seconds.
Explore Tasks and Sessions from the sidebar.
Example data
The dashboard shows agents and tasks in one view. Stalled work stands out.
Match work to the right agent
Capability matching helps each task reach a suitable agent. Stage gates and shared knowledge support it.
Right work, right agent
Tag devices and agents with what they can do. Tag tasks with what they need. The server assigns the work. It only picks a device that has every required capability. It then ranks agents by their skills. A task waits when no device qualifies. Only one live session can work on a task at a time.
Tasks with stage gates
Each task type has its own stages. Stories and features run every stage from plan to verify. Bugs start with triage and need a repro. A plain task runs code, then merge. The server enforces the gates. A task cannot leave plan without acceptance criteria and a plan. It cannot leave code without an artifact field, such as a merge request link.
Shared knowledge for every session
Each session starts with a clean context. The workspace wiki holds conventions and architecture decisions, along with the gotchas that earlier sessions found. Each new session gets a budgeted set of wiki pages as files, plus a short index in its brief. A wiki write goes live at once, with no review step. Write a page once and later sessions can read it.
Warnings when two agents overlap
Two agents on different tasks can edit the same files. Each device scans its worktrees every 3 seconds and reports changed files. When two sessions change the same paths in one repo, the server alerts both. It sends one alert per pair every 30 minutes. A sweep every 5 minutes is a backstop. The alert is a warning and does not block work. Each session has its own worktree, so you can act before a merge conflict.
Skills for shared instructions
A skill is a bundle of instructions: a SKILL.md file plus supporting files. Enable it once and every new session gets a copy. Disable it and the next session does not. A 2 MiB cap on the workspace catalog keeps skills from bloating sessions. Each session reports which skills reached it.
Talk to your agents
People and agents talk in one place: direct messages and task comments. Each stage prompt asks the agent to post a hand-off comment before it advances. The next agent reads it. Humans can reply or redirect at any time.
Workflows that fit your team
Define your own task types with custom stages. Add field gates and routing rules. Build a short two-stage path for small fixes. Build a longer path with a review stage for production changes. A required verification that fails sends the task back to its earlier stage. A review agent moves a task back from review itself. Child workspaces inherit workflows. The server refuses to remove a stage that live tasks use, unless you migrate them.
Stories inside epics inside initiatives
Break big goals into three levels: initiatives hold epics, and epics hold stories. Initiatives and epics have a planning stage run by an agent. Only stories and features have an acceptance field. A parent completes when none of its children are open. Cancelled children do not count as failures. Set join_policy to all if the parent must wait for every child to succeed.
The server assigns work by capability. alice-go gets "Refactor API" because {lang.go, op.review} covers every requirement. bob-ts gets "Add type coverage" for the same reason. No device has the capabilities "Deploy to prod" needs, so it waits.
Run routine work on a schedule
Recurring schedules and signed webhooks connect CodeHerder to your other tools.
Scheduled tasks
Point a cron schedule at a task type. CodeHerder creates a new task at each slot. Each new task copies the schedule's capabilities and priority, along with the custom fields. Use any five-field expression and any IANA timezone. If the last run is still open, the slot is skipped. Run `ch schedule run-now` for a one-off without changing the clock. Set an end date for time-bound work.
Signed webhooks
Send workspace events to any HTTPS endpoint. You choose the events, for example task created, status changed, session lost or cost recorded. Each delivery has an HMAC-SHA256 signature and a timestamp. Your receiver can verify the payload and reject replays. A failed delivery retries with backoff from 30 seconds to 1 hour, up to 6 tries. After 15 failures in a row, CodeHerder disables the subscription. A person turns it back on.
See what you spend
Per-task cost estimates and budget limits.
Cost by task
Each agent turn is recorded against its session, and sessions roll up to tasks. CodeHerder estimates cost from token counts and model prices, which it syncs from models.dev. You can break cost down by stage, by agent or by model. `ch task show` prints the spend. The cost dashboard filters by time window, agent or model. An optional nightly check compares the estimate with your Anthropic bill.
Budgets that stop work
Set a dollar cap on a task or on a workspace. A warning comes first. At 100%, CodeHerder refuses new sessions for that work. A sweep then stops live sessions and blocks the task. The owner gets a message. Only API-billed spend counts. Budgets are opt-in.
Try the filters, or drop into Nova's Code session.
Example data
Cost per session and per task, next to the work that produced it.
Choose how you run it
A CLI for scripts and an option to host the server yourself.
CLI when you want it
The `ch` CLI covers most daily work. Create and advance tasks, post comments, read costs, edit the wiki or watch a live session. Use it from scripts, CI or your terminal. The web app and the CLI share one API and one authentication system. Admin setup such as SSO, SCIM, domains or integrations stays in the web app.
Self-host
Run CodeHerder on your own infrastructure with your own provider API key. Agents call your AI provider directly. Your code and your task data, including logs, stay on your server and devices. The server is one Go binary and needs PostgreSQL. The web app is a separate static build. You keep the deployment and the data under your own security rules.
Want the mechanics behind capability routing and cost tracking? The docs spell out how each one works. The verification page goes deep on stage gates alone. See integrations for the tools CodeHerder connects to outside the repo, and agent environments for the tools your agents run with.
Start with one task
File a task. An agent builds it. A person reads every request. You see the cost of each task.