The platform ships zero AI. Instead, it hands your request queue to the agent you already run. Client asks, your agent builds, you approve the preview. That is the whole loop.
Run the listen daemon for full autonomy, or point any MCP-capable agent at the workspace and drive it yourself. Same queue, same tokens, same audit trail.
Create a token with a device-code flow, scope it to your workspace, revoke it anytime. Your agent authenticates like a first-class citizen, not a scraped session.
A long-running process on your machine that polls for queued requests on auto-enabled projects and spawns your agent per request. Claude Code by default; point it at any binary.
An MCP server exposes requests, clients, skills, and quotes to your agent: list the queue, mark in progress, attach a branch, read the skill for the job.
The platform serves a library of Markdown skills your agent reads before working: how to build, edit, and operate on client sites. Consistency without a model in the loop.
Branch, PR, duration, and outcome logged per run on the Agents & MCP screen. If a request fails three times, it flags for a human instead of looping.
A request lands in the queue from the portal or a pinpoint on the live site.
Your agent gets the request, the client context, and the matching skill.
The agent works in that client's site repo and pushes a branch, never to production.
A preview deploy attaches to the request. The reviewing side gets a link, not a diff.
You or your client approve. Merge ships it. The run is logged with the branch and PR.
Platforms that bolt a model inside choose your tools for you and charge you per thought. Agency Label stays the system of record: the queue, the portal, the approvals, the billing. The intelligence is yours: your agent, your API keys, your machine, upgraded whenever you like without waiting on us. That is also why agent runs are unlimited on every tier. They run on your side, so we have no reason to meter them.
No AI features. No AI fees. The platform never calls a model. "Connect your agent" means exactly that: your agent, your credentials, our queue and approval gates around it.
Works with your stack. Claude Code out of the box. Anything that speaks MCP or can run as a binary works: the daemon does not care what does the thinking.
Create a workspace, add a client, run one command. The first request your agent ships changes how you scope your week.