How Build Station works

Build Station bridges the Meshly Build dashboard (where tasks live) and the AI agent (Claude Code) that actually does the work on your machine. Understanding the flow helps when things go sideways.

The flow

flowchart LR
  A[Dashboard task created] --> B[Backend dispatches to agent]
  B --> C{Build Station online?}
  C -- yes --> D[SSE push]
  C -- no --> E[Queued — picked up on next poll]
  D --> F[Claude Code spawned in working dir]
  E --> F
  F --> G[Agent edits code, runs tests, calls MCP tools]
  G --> H[Activity + completion posted back to dashboard]
  H --> I[Task in Review]

Key concepts

Project. A repository on the Meshly Build dashboard mapped to a local working directory on this machine. Tasks live inside a project. An agent can only work in a project if Build Station knows where the local checkout is.

Agent. An identity that does work — typically tied to a role (developer, infra, reviewer, etc.). Each agent has its own credentials, its own task queue, and its own model preferences. On Build Station you enable per-agent on each project.

Task. A unit of work assigned to an agent. Tasks flow Backlog → To Do → In Progress → Review → Done. Build Station only picks up tasks in To Do state assigned to an agent enabled on this machine.

Working directory. The local folder Build Station opens for the agent. The agent has full filesystem access there — it can read, write, run git, run tests, run shell commands.

Two ways tasks reach Build Station

• Live (SSE): When Build Station is open and connected, tasks push to it in real time. Spawn latency is sub-second.

• Polled (every 30s): Tasks created while Build Station was offline get picked up on the next 30-second poll after it comes back online.

What happens inside a Claude session

Build Station doesn't write the agent's prompt itself — the dashboard does. When the agent picks up a task it receives a prompt that contains:

• The task title and description (your words to the agent).
• Project context: conventions, recent decisions, similar past work.
• Role guidance for whatever role the agent has (developer / infra / reviewer).
• The list of MCP tools it can call.
• Strict workflow rules: track via tasks, commit before completing, no fakes, etc.

Claude then works in the project's directory, reading files, editing code, running commands, and calling MCP tools to update the dashboard.

Note: The agent uses your Claude subscription

Build Station launches the claude CLI you have installed locally. Whatever account that CLI is signed into — your subscription, an API key, whatever — is what gets used. No API credits are spent by Build Station itself; it's a thin orchestrator.

When Build Station is closed

Tasks queue server-side. Nothing is lost. When you next launch Build Station:

1. It syncs project + agent state with the server.
2. The queue poller starts and finds any tasks that arrived while you were offline.
3. They spawn in order, one at a time per project.

If you keep Build Station closed for a long time the queue can pile up. There's nothing automatic that drops old work — you just see a flurry of spawns when you next launch.