Running a job search with Claude Code and Codex, and what we learned building the MCP server behind it
A lot of people now let an AI agent do the boring parts of a job search: finding postings, reading them, judging fit, filling in forms. Claude Code and Codex are surprisingly good at this. What they are bad at is remembe
A lot of people now let an AI agent do the boring parts of a job search: finding postings, reading
them, judging fit, filling in forms. Claude Code and Codex are surprisingly good at this. What they
are bad at is remembering. Close the session and the agent forgets which jobs it read, which ones it
applied to and what it wrote in each form.
We built ApplyKoala to be that memory: a job application tracker with a
remote MCP server, so the agent keeps the whole search in one place and you can see it on a board.
This post is how the setup works and a few things we learned along the way.
The setup in five minutes
Claude Code:
claude mcp add --scope user --transport http applykoala https://applykoala.com/mcp
Then type /mcp and sign in. Codex, Cursor, Claude on the web and ChatGPT take the same address;
the MCP page has the one-liner for each.
Then talk to it like a person:
Here is my resume. Find full stack jobs in Singapore that fit me, read each posting in full,
research the company, and save the ones worth applying to with your honest view of the fit.
The agent searches, calls save_job with the posting copied word for word, add_company_research
with a source link for every fact, and set_job_fit with quotes from your resume. When you apply,
record_application keeps what was sent: the resume version, the cover letter, every form answer.
Rules we put in the tools, not in a system prompt
Agents are eager. Left alone they summarise postings, guess salaries and invent experience. We
ended up enforcing the important rules on the server, where the agent cannot talk its way around
them:
- Quotes must exist. A fit verdict has to quote the user's own resume or profile for each requirement they meet. The server checks every quote against the saved text and rejects the call if a quote is not really there.
- Every research item needs a source URL and a date. No source, no save.
- Nothing is deleted for good. Earlier versions of each record are kept for 30 days, so an over-eager agent can be undone.
- ApplyKoala never submits an application. Its tools find, save and track. If the agent fills in a form in its own browser, the user decides what gets sent.
- Saved web text is data, not instructions. Postings and emails are returned in a way that tells the agent not to follow anything written inside them.
Short tool descriptions with one rule each worked better than a long instruction block. Smaller
models in particular follow "Copy the posting verbatim into description" when it sits right on the
save_job field.
Finding jobs before everyone else
The part people ask about most is job alerts. Instead of scraping job sites, we read the public job
board APIs that companies use for their own careers pages (Greenhouse, Lever and Ashby), for about
9,500 companies. An hourly EventBridge rule fans out to Lambda workers through SQS, diffs each board
against the last snapshot in DynamoDB, and pushes new and closed jobs to a Cloudflare Worker. Users'
alert rules only filter that shared pool, so a thousand subscribers cost the same crawl as one.
The agent reads matches with get_new_jobs, which gives each job's original posting time. That
timing turned out to matter. With more people applying through agents, a posting can collect a lot
of applications in its first hours. So we:
- store the exact posting time with each saved job,
- added a "within 6 hours" bucket to the dashboard's Does applying early help? card, so users can see their own interview rate by how fast they applied,
- and chart when companies actually post. For Singapore over the last two months: mostly Tuesday to Friday, 9 am to 6 pm local time, peaking around 4 pm, with almost nothing at weekends.
If you want the agent to check every hour without you, Claude Code's cloud scheduled tasks can run
a prompt hourly with your connectors. We wrote up the prompt and the setup for Claude Code, Claude
Desktop, Codex and ChatGPT here:
Have your AI check for new jobs every hour.
Stack, briefly
- Web app: Preact, local-first in IndexedDB; optional sync stores an encrypted copy (a key per user) on Cloudflare Workers, Durable Objects and D1.
- MCP server: remote, streamable HTTP, OAuth 2.1 with PKCE and dynamic client registration. 47 tools.
- Crawler: AWS Lambda, SQS, DynamoDB and EventBridge, all in Terraform.
The tracker is free in the browser; connecting an AI needs sync (US$3.99 a month, 7-day free trial,
no card). The Claude Code plugin is public at
github.com/applykoala/applykoala-plugin.
Questions about how any of this works are welcome in the comments.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.