The dashboard
The dashboard is a local web page served by agentd dash. Open it once you
have connected your coding agents.
agentd user add YOUR_NAME --role owner # once
export AGENTD_USER=YOUR_NAME # add it to your shell profile
agentd dash --serve 127.0.0.1:21480
agentd dash link --user YOUR_NAME
AGENTD_USER tells AgentD who starts and pays for the workers you start from a
terminal. agentd run refuses a user it could only infer from your operating
system account.
Open the printed link in your browser. It works once, within 2 minutes, and
signs that browser in for 12 hours. If you open the dashboard from another
machine, add --base with the address that browser uses, for example
agentd dash link --user YOUR_NAME --base http://HOST:21480.
Users have a role: owner, dev or viewer. Add more people with
agentd user add NAME --role dev.
Views
| View | What it shows |
|---|---|
| Inbox | What workers need from you: workers blocked on input, failed workers, and approvals to answer. The dashboard opens on it. |
| Issues | Tasks and their results, with the linked pull requests. Merge a linked PR from here. |
| Team | Connected sessions and running workers. Open a session's console to see its live output. |
| Chat | A conversation with a role, which can propose worker tasks. |
| Insights | Includes History: finished tasks, newest first. |
Chat
Chat answers with a model on an OpenAI-compatible gateway, set under
[gateway] in ~/.local/share/agentd/config.toml. It is never answered by a
coding agent's transcript. To use the LibraOS model API, see
Models and billing. agentd doctor reports whether the
gateway is reachable and how many models it lists.
Roles. A chat is with a role. Two roles ship with AgentD: assistant,
which can search, read progress and file coding work for the implementer
role, and writer, which can only search. To add your own planner, create
agents/NAME.md in AgentD's root (~/.local/share/agentd/agents/) with front
matter such as:
---
chat: true
model: anthropic/claude-opus-5-5
description: Splits a request into tasks, one per worker.
---
Plan the change as tasks a worker can finish and check on its own.
chat: true puts the role in the Chat pane, and a chat role needs model, the
gateway's model name. tools lists what its model may call: kb_search,
ledger_progress, handoff_offer and dispatch, all four by default, or
tools: none. agentd roles validate checks role files.
Starting workers from Chat. Chat proposes tasks only when the dashboard may dispatch:
[dash]
dispatch = true
or agentd config set dash.dispatch true. The planner then proposes one task
per worker, each with its goal, acceptance criteria, checkout, worker, budget
and network limits. Nothing starts until a person with write access chooses
Review and start, then Confirm and start. Conversation text and the
model's own tool calls cannot confirm a task. A budget is required: an explicit
budget, then the role's, then 30 minutes.
Chats are personal. On a ledger with more than one user, a chat belongs to its person, and viewers can read their own but not write.
Next: Run and review work.