jobhound 0.16.0

Action-based CLI for tracking a job hunt

jobhound

Action-based CLI for tracking a personal job hunt. Status changes are a consequence of recorded events (apply, log, withdraw, …), not direct field edits.

Install

uv tool install jobhound   # or: pipx install jobhound

Exposes the jh command.

For tab-completion, run:

jh completion install

See jh completion --help for per-shell options.

Usage

jh new --company Acme --role "Senior Engineer"
jh apply acme
jh log acme --channel email --direction to --who recruiter \
            --body draft.md --next-status screen
jh note add acme "Recruiter mentioned a hybrid setup"
jh withdraw acme
jh list
jh show acme                # human-readable detail; --json for the envelope
jh export --active-only     # bulk JSON envelope to stdout

Each command is a verb on a single opportunity. Run jh --help for the full set: new, apply, log, withdraw, ghost, accept, decline, note, priority, tag, link, contact, list, edit, archive, delete, sync, show, export.

jh export filters: --status and --priority (comma-separated or repeatable), --slug (substring), --active-only, --include-archived.

Adding a job from a URL

Create an opportunity by scraping a job posting instead of typing the details:

jh new --url https://www.linkedin.com/jobs/view/4383908452

jobhound extracts the company, role, location, and description, stores the posting as a posting link, and saves the full text as job-description.md. Extraction runs locally with no LLM: LinkedIn is read from its page metadata, and other job sites are read from the schema.org JobPosting data most publish for Google Jobs. Tracking parameters in the URL are ignored — deduplication keys on the posting's canonical URL, so the same job added twice is rejected.

Most postings are fetched without logging in. For a posting behind a login, let jobhound reuse the session from your browser — grant permission once:

jh config set allow-browser-cookie-access true

jobhound then reads the cookies for that site from your default browser (configurable with jh config set cookie-browser <name> and cookie-browser-profile) and replays them to fetch the posting. Cookies are read only for the target site, used for that fetch, and never stored.

Managing opportunity files

jh file provides uniform access to files inside an opportunity (CVs, research notes, correspondence, etc.). Available subcommands:

• jh file list <slug> — list files
• jh file show <slug> <name> — print to stdout; use --out <path> to export
• jh file write <slug> <name> --content <str> (or --from <path>)
• jh file append <slug> <name> --content <str> (or --from <path>)
• jh file delete <slug> <name> --yes

Pass --base-revision <r> (the revision string from a prior jh file show or jh file write) to enable optimistic-concurrency conflict detection. For text files, a 3-way merge is attempted automatically; binary conflicts surface the current file's metadata and suggest an alternate name.

Storage

Per-opportunity data is stored under $XDG_DATA_HOME/jh/ (defaults to ~/.local/share/jh/) as a TOML file plus markdown notes and correspondence. The data root is a git repo with auto-commits on every state change — your history is auditable and you can push it anywhere.

AI integration (MCP)

jh ships a Model Context Protocol server so AI clients (Claude Desktop, Claude Code, Continue, Zed, …) can read and modify your job hunt directly. The MCP tools cover read operations (list, show, stats, files, file content), creating an opportunity from a job-posting URL (create_from_url), state transitions (apply, log, withdraw, ghost, accept, decline), field setters, relation operations (tags, contacts, links), opportunity ops (notes, archive, delete), and a uniform file API (read, write, import, export, append, delete) with optimistic-concurrency conflict detection and 3-way merge for text files.

Install the optional extra:

uv tool install 'jobhound[mcp]'

Then point your MCP client at the server. For Claude Desktop, add to claude_desktop_config.json:

{
  "mcpServers": {
    "jobhound": {
      "command": "jh",
      "args": ["mcp"]
    }
  }
}

For zero-install discovery (no uv tool install needed):

{
  "mcpServers": {
    "jobhound": {
      "command": "uvx",
      "args": ["--from", "jobhound[mcp]", "jh-mcp"]
    }
  }
}

The same pattern works for Claude Code (.mcp.json), Continue, Zed, and any other MCP-spec-compliant client.

By default the AI gets full CLI parity — including writes. Most MCP clients show each tool call to the user before executing it; that's the consent layer. The one tool that requires explicit double-confirmation is delete_opportunity, which needs confirm=true in the call args (otherwise it returns a preview only, no side effects).

Status

Pre-1.0. The CLI surface is stable; semantic-versioned releases via Conventional Commits.