Skip to main content
Glama
Hunt

Hunt

by meetshukla

Hunt

Self-hosted AI job search for Claude Code, Codex and any MCP agent.

GitHub stars License: MIT PRs welcome

Hunt collects new postings from the job boards of the companies you pick, keeps the ones that pass your filters, and tracks every application. Your agent reads the new jobs over MCP, judges the fit against your resume, applies, and records what happened. It runs in your own Cloudflare account, on the free plan.

Deploy to Cloudflare

If Hunt helps your job search, star the repo. Stars help other job seekers find it.

What it does

Hunt reads companies' own job boards, the Greenhouse, Lever, Ashby and Workday pages behind their careers sites. A posting reaches you about an hour after it goes up, straight from the company, with its full description.

The app does the steady work and your agent does the thinking. There is no AI inside and no API key to pay for.

Hunt

Your agent

Find

Reads company job boards, remote job boards and RSS feeds, each about once an hour

Picks the companies and boards that fit you

Filter

Keeps jobs that pass your filters: titles, places, remote regions, how recent, words to skip

Sets the filters with you

Judge

Reads each new job against your profile and writes a fit score with a one-line reason

Apply

Writes the tailored resume and applies on the company's site

Track

Status, notes, contacts and a timeline for every job

Records each step as it goes

flowchart LR
  boards["Company job boards<br>Greenhouse, Lever, Ashby,<br>Workday and more"] --> app
  remote["Remote job boards<br>and RSS feeds"] --> app
  app["Hunt<br>in your Cloudflare account"] <-->|MCP| agent["Your agent<br>Claude Code, Codex"]
  app <--> web["The website<br>on your laptop and phone"]
  agent --> apply["Applies on the<br>company's site"]

Related MCP server: job-agent-mcp

Get started

You need a Cloudflare account (the free plan is enough) and an agent that speaks MCP.

1. Deploy

Click Deploy to Cloudflare above. Cloudflare copies this repo to your GitHub, builds it, and asks for HUNT_TOKEN. The token is a long random password, and openssl rand -hex 24 makes a good one. Keep it in your password manager.

Or deploy from your terminal with Node 22.18 or later:

git clone https://github.com/meetshukla/JobSearch-MCP hunt
cd hunt
npm install
npx wrangler login
npm run deploy                              # prints https://hunt.<you>.workers.dev
openssl rand -hex 24                        # your token
npx wrangler secret put HUNT_TOKEN          # paste the token

Open the URL and sign in with the token. To use your own domain, run npm run build && npx wrangler deploy --domain jobs.example.com.

2. Connect your agent

The MCP server is <your url>/mcp. It takes your token as a bearer token. Put the token in your shell profile first:

export HUNT_TOKEN=...

Then add the server to your agent.

# Claude Code
claude mcp add --transport http hunt https://hunt.<you>.workers.dev/mcp \
  --header "Authorization: Bearer $HUNT_TOKEN"

# Codex
codex mcp add hunt --url https://hunt.<you>.workers.dev/mcp \
  --bearer-token-env-var HUNT_TOKEN

Other clients take this JSON:

{
  "mcpServers": {
    "hunt": {
      "type": "http",
      "url": "https://hunt.<you>.workers.dev/mcp",
      "headers": { "Authorization": "Bearer ${HUNT_TOKEN}" }
    }
  }
}

The website's Settings > Connect agent page shows the same commands with your URL filled in.

3. Set it up with your agent

Open your agent in this repo and ask:

Read AGENTS.md and set up Hunt for me.

AGENTS.md walks the agent through the setup. It asks for your resume, where you can work, the titles you want and your deal-breakers. Then it saves your profile, sets the filters and picks the companies to follow. The first jobs arrive within the hour.

You can also do it yourself on the website, under Settings.

Use it every day

Talk to your agent the way you would to a friend who helps with your search.

You say

Your agent

"Review my new jobs."

Reads the jobs it hasn't scored yet, gives each a fit score and a reason, and skips the ones that don't fit

"Which ones should I apply to this week?"

Lists the best-scored open jobs with the reasons

"Apply to the Supabase one with a tailored resume."

Checks the posting is still open, writes the resume, applies, and marks the job Applied

"Who should I write to about it?"

Finds the hiring manager and recruiter with the find-job-contacts skill and saves them on the job

"Reddit rejected me."

Moves the job to Rejected

Your agent reads only the Jobs view, which holds new jobs. Anything you applied to or skipped leaves it, and unscored_only hides the jobs your agent already scored, so it reads each job once.

The website

The website shows what your agent sees, and you can do everything there too. It has three views.

View

Holds

Jobs

New jobs that pass your filters. Scored jobs first, then by the day they were posted.

Applied

Jobs you applied to, grouped by In process, Applied and Rejected

Skipped

Jobs you or your agent passed on

It works on your phone as well.

Keyboard shortcuts cover everything. Press ? in the app for the list. The most used are j and k to move through jobs, s to change a status, o to open the posting and / to search.

Tools

The MCP server has nine tools. The website calls the same ones, so you and your agent always see the same data. The definitions are in server/tools.ts.

Tool

What it does

get_settings

Returns your profile, your filters and what setup still needs

update_settings

Saves your profile (resume and what you want, as plain text) and your filters. With preview, it counts the jobs a filter change would show.

add_sources

Adds company names, careers pages, job board URLs, RSS feeds or a starter list

list_sources

Lists sources with their state, errors and last read time

update_sources

Pauses, resumes, removes or re-reads sources

list_jobs

Lists one view (jobs, applied or skipped) as short rows. unscored_only returns the jobs your agent hasn't judged yet.

get_job

Returns one job with its full description, contacts and timeline

add_job

Tracks any job from a link, whatever your filters say

update_job

Sets the fit score and reason, the status, a note or contacts

A job's status is one of new, skipped, applied, in_process and rejected.

Sources

  • Company job boards. Greenhouse, Lever, Ashby, SmartRecruiters, Workable, Recruitee and Workday. Add a company by name and Hunt finds its board, or paste its careers page or board URL.

  • Remote job boards. Remotive, Himalayas, Jobicy, We Work Remotely and Arbeitnow, on by default. Remove the ones you don't want.

  • RSS and Atom feeds of job postings.

  • One job at a time from any posting page with schema.org job data, through add_job.

  • A starter list of 1,000 tech companies with public job boards, mostly in the US and Canada, in lists/north-america-tech.json. It is optional.

LinkedIn and Indeed have no public job feeds, so Hunt doesn't read them. Most companies that post there also post on their own board, which Hunt reads. For the rest, give your agent the link and it tracks the job with add_job.

Filters

Hunt has opinions, and every one of them is a setting you can change or turn off.

Filter

Default

Titles to keep

software engineer, software developer

Titles to skip

intern, internship, co-op, new grad, sales engineer, solutions engineer, support engineer

On-site and hybrid places

Anywhere

Remote regions

Any region

Remote jobs that don't say where

Kept

Posted within

7 days

Jobs with no posting date

Hidden

Words in the description that rule a job out

None

Companies to skip

None

Phrases match whole words in any case, so "engineer" doesn't match "engineering". AGENTS.md explains each filter for your agent.

Skills

Two optional skills in skills/ give your agent a routine. Copy a folder into your agent's skills folder, for example ~/.claude/skills/.

  • review-jobs scores new jobs against your profile the same way every day and brings the best ones to you.

  • find-job-contacts finds the hiring managers and recruiters for a job, with LinkedIn links and verified emails, and saves them on the job. It uses treg by default, and any people-data provider works.

Limits and cost

Hunt runs on Cloudflare's free plan, so it costs nothing. One Worker and one Durable Object hold everything, with your data in the Durable Object's SQLite database.

On the free plan, each run of a Worker can make at most 50 requests to other websites, and reading one job board takes one request. So Hunt can't read 1,000 boards in a single run. It runs every 15 seconds instead, and each run reads four company boards, one remote board or feed, and six job descriptions. That is 240 runs an hour, enough to read about 1,000 company boards an hour. With more boards, each one is read less often.

FAQ

Does it apply to jobs for me? No. Your agent applies, in its own browser, after you agree. Hunt keeps the list and the record.

Does it need an OpenAI or Anthropic key? No. Hunt runs no AI model. Your agent judges the fit with its own model.

Where is my data? In your Cloudflare account, in one SQLite database. Nothing goes anywhere else.

Can several people share one deployment? No. A deployment is for one person: one profile, one token. Each person deploys their own, which is free.

Does it work outside North America? Yes. Places and remote regions are free text, and you choose the companies. The starter list is the only part that leans North American, and a list for your region is a welcome pull request.

Develop

npm run dev                                          # http://localhost:8787, no sign-in
node scripts/mcp.ts http://localhost:8787            # list the tools
node scripts/mcp.ts http://localhost:8787 list_jobs  # call one
npm test                                             # filter tests
npm run typecheck

Path

Holds

server/filters.ts

The filter, one pure function used when collecting and when listing

server/lib/collect.ts

The collector that runs every 15 seconds

server/lib/sources.ts

Adding and resolving sources

server/lib/jobs.ts

Views, one job, and tracking

server/sources/

Readers for each job board system, remote board and RSS

server/tools.ts, server/mcp.ts

The tools and the MCP server

shared/types.ts

Shapes shared by the server and the website

client/

The website, in React. client/DESIGN.md has its design rules.

worker/

The Cloudflare Worker and the Durable Object

Contributing

Pull requests are welcome, from a one-line fix to a new job board system. CONTRIBUTING.md shows how to run Hunt on your machine and what to check before you open one. These help the most:

Not sure where to start? Open an issue with what you want to change, and we'll work out the shape together.

If Hunt found you a job, or just saved you time, a star helps other job seekers find it.

License

MIT

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to scrape job offers and filter them based on user-defined criteria, using MCP resources, prompts, and tools.
      -
    • A
      license
      Not graded
      quality
      D
      maintenance
      MCP server that aggregates and deduplicates job listings from multiple public sources, ranks them against a user's resume, and exposes tools for searching, viewing details, explaining fit, and tracking applications.
      MIT
    • F
      license
      A
      quality
      D
      maintenance
      Enables searching job listings, tracking applications, managing resumes, and tailoring resumes to job posts, all locally via MCP.
      20
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables an MCP-capable coding agent to run a local-first job search: recording verified opportunities in a private SQLite tracker, tracking follow-ups and outcomes, and preparing evidence-based resumes, application answers, and outreach drafts. It keeps all personal data on the user's machine and stops before submitting applications, uploading documents, or contacting anyone.
      1
      MIT