Apply to jobs with Codex CLI, even on Workday

3 September 2026 · 8 min read · by Akhil Gorantala

Workday applications are the final boss of job hunting: six pages, an account per company, and a form that re-asks everything your resume already says. OpenAI's Codex CLI plus Dowser turns that slog into a supervised walk: Codex fills each page from your profile, you review and advance. Here is the setup and the page-by-page pattern.

Step 1: setup with Codex CLI

Either paste the agent-guide line and let Codex do everything:

Set up the Dowser browser extension for me: fetch https://dowser.click/install.md and follow the steps in it, then confirm Dowser is connected.

Or do the two steps by hand: download and pair from dowser.click/install.md, then register the server: codex mcp add dowser -- node /Users/you/dowser/mcp/server.mjs (absolute path). Add the extension to Chrome when asked, and confirm with "list my open tabs".

Step 2: your profile, once

Give Codex the facts it will be asked for on every page: contact details, links, work authorization, veteran/disability self-identification preferences (tell it "decline to answer" if that is your choice, it is a normal option), education and dates, and the standing rules: show every draft, never click Submit or Next without being asked, stop at anything that looks like an account creation or password field.

Step 3: the Workday page-by-page pattern

• Account wall: you create the account and sign in yourself; password fields are refused by rule. Then hand back.
• My Information: fill this page from my profile, then stop. Codex fills a dozen fields in about a second of wall time, you eyeball, you click Next.
• Experience: Workday's repeating job-history blocks go one block per instruction: "add my most recent role from the profile". Dropdown months and years select by label.
• Application questions: factual ones fill from profile; judgment ones get drafts in chat first.
• Review page: yours entirely. Codex can dowser_read it back to you as a checklist, but the Submit click is you.

Six pages, maybe eight instructions, and the only typing you did was your password. The same pattern collapses to one instruction on single-page boards like Lever and Greenhouse.

What stays manual

• Account creation, sign-in, passwords: refused fields, deliberately.
• Resume file upload: file inputs are unsupported; drag it yourself.
• CAPTCHAs: yours.
• Next/Submit on each page: Codex is instructed to stop, and destructive clicks require confirmation anyway.

Token usage, honestly

A six-page Workday run through a screenshot agent is dozens of vision round-trips, one per field or near it. Through Dowser it is: read the posting once, read each page's outline once, and sentence-sized fill calls that cost effectively nothing. Whole-application context stays smaller than one screenshot, which matters if you are doing ten of these in an evening on a metered plan. Numbers and sources: the token-math post.

FAQ

Does Codex CLI support MCP like Claude Code does?

Yes: codex mcp add registers stdio servers; Dowser is one. The tools appear as dowser_find, dowser_act, dowser_read, dowser_navigate, dowser_tabs.

Workday looks different per company; does finding break?

Fields are matched by meaning against the accessibility tree, not by pixel position or exact markup, so the usual per-company Workday reskins do not matter. Genuinely custom widgets can need a clarifying instruction.

Can it reuse answers between companies?

Yes, that is the point of the profile message: the facts persist in the session and every page fills from the same source. For long-term reuse, keep a profile file and have Codex read it.

What if a fill lands in the wrong field?

Low-confidence matches return candidates instead of filling, and every action is visible under the cursor; if something lands wrong anyway, say "clear that and put it in the other field" and it is corrected in milliseconds.

Other assistants?

Same workflow: Claude Code, Cursor, Gemini CLI, Windsurf.