Why we taught Atlas to speak MCP
When we made Atlas headless for agentic use, the deal was simple: hand a constrained agent account a token and let it drive your own search through the same typed API the dashboard uses. It worked, but it left one cost on the table — every agent needed bespoke glue code to speak our API. That cost is gone. Atlas now runs a first-party MCP server, so any client that speaks the Model Context Protocol can operate your account with zero integration work.
MCP is the plumbing standard the agent ecosystem has settled on: a client asks a server what tools exist, then calls them. Instead of teaching your agent about Atlas, you point it at one endpoint and it discovers the surface itself.
What it actually is
The MCP server is a thin shell over the engine you already use. Every tool wraps one of the same typed procedures the dashboard calls — no parallel code path, no special agent logic, no back door. If the dashboard can do it, the wrapper does exactly that; if the dashboard can't, neither can your agent.
It is also deliberately not a platform. The server operates one account: yours. There is no third-party hosting, no shared workspace, no public developer API. An agent with a token is a constrained version of you, nothing more.
What an agent can do today
The surface is nineteen tools across three domains, and each one maps to a capability you already know from the app.
- Run your daily search and read the ranked, scored shortlist
- Move jobs across pipeline stages and keep your board current
- Add notes to listings and read note history
- Read dashboard stats, stage history, and job locations
- List your tracks and archives
- Check search status, history, and how much quota remains
Bounded by design
Every limit that applies to you applies to your agent, because the agent is your account. One search per day. No overnight digest email — there is no human inbox behind an agent. Tokens are issued by the admin, hashed at rest, shown once, revocable instantly, and can now carry an expiry you choose at issuance: thirty, sixty, or ninety days, or none at all.
That bounded shape matters more as agents get more capable. The market is full of tools that hand automation the keys and hope for the best — a pattern we've written about in how AI is rewriting the rules of the job search. Atlas takes the opposite bet: the agent gets a narrow, observable lane, and you keep the steering wheel.
How to get access
Agent accounts are issued on request while the capability matures. Ask for one, get a token, and drop the endpoint into any MCP client's configuration — the exact snippet lives on the agents page. From request to a working agent is a few minutes, most of which is your client's config file.