Give your agent a job search it can actually run.
One MCP server and twenty-one tools, wired to the same apply engine the app runs on. Connect Claude Code, Cursor, Codex or anything else that speaks MCP, and it can search your matches, tailor a résumé and file the application without you leaving the chat.
Your agent already knows how to talk. This gives it something to do.
Most job tooling stops at search results, which leaves the model to write the application and you to paste it somewhere. This exposes the whole pipeline — the feed, the profile, the tailoring, the submit and the tracker — as tools an agent can call, with the parts that matter still done here rather than in the prompt.
One address, and your client already knows the rest
A standard remote server, so there is no per-client build and nothing to keep updated. One command wires up whichever clients you have installed, hands each the same URL, and the browser handles the sign-in.
Twenty-one tools, not a search endpoint
Jobs, applications, profile, applying and documents. Your agent can read the feed, save a role, check whether you are eligible for it, file the application, follow the run, and update the tracker afterwards.
The two that spend something say so
Reads are marked read-only, so most clients run them without stopping to ask. Filing an application is marked destructive, because it puts your name on a real form at a real company, and your client should interrupt you before that happens.
It asks you when it gets stuck
If a run hits a screening question the engine will not guess at, it comes back through your client and waits for the answer, then finishes the submission. Nothing is invented to keep the run moving.
The agent steers. It does not write.
Tailored résumés and cover letters are produced here, from your actual résumé and the actual posting, and checked before they go anywhere. The connected model decides what to apply to; it never writes the document that gets submitted.
Nine permissions, and an off switch
A client that asks for nothing gets read-only. Anything more has to be requested by name, and the approval screen lists exactly what was asked for before you allow it. Revoke any agent from Settings whenever you want.
The parts you will want before you wire it up
Enough to know whether it fits your setup without installing anything first.
- One command, or one line of config
- `npx @aiesphere/mcp-connect` detects the clients on your machine and writes each one’s config, backing up anything it touches. Prefer to do it yourself? It is a single URL in your client’s `mcpServers` block.
- Stdio clients are covered too
- Anything that only speaks stdio, or does not run OAuth itself, goes through the standard `mcp-remote` bridge — same server, same URL, one extra line.
- Your state, pinnable
- Three resources — your profile, your résumés and your application board — so a long conversation can hold your situation instead of fetching it again every turn.
- Four prompts, as slash commands
- Apply to your best matches, tailor for one job, review the pipeline, or unstick a run that stalled. They show up in most clients as commands rather than something to remember.
- Follow a run, or fire and forget
- Filing an application returns a run id straight away. Ask it to wait instead and it streams progress while the form is filled, so the agent can report what happened rather than guess.
- It is the same account
- Everything an agent files lands in the same tracker as everything else, against the same profile and the same résumés. There is no separate agent-shaped copy of your job search.
What it will not do
The useful half of handing an agent your job search is knowing where its reach stops.
It will not apply to anything without the submit permission, and that one has to be asked for by name.
It will not invent an answer to a screening question. If it cannot ground one, it stops and asks you.
It will not let the connected model write your résumé or your cover letter. It can ask for one; it cannot author it.
It will not act on a company that has taken the posting down, and it will tell you that is what happened.
It will not survive a revoke. Cut an agent off in Settings and its next call fails, mid-conversation.
Point your agent at it.
One command, whichever client you already use. Sign in once in the browser and it is connected.