amt

amt is a personal job-search toolkit: a CLI plus an MCP server (amt-mcp) that puts a clerk — der Sachbearbeiter — inside your AI coding agent. It crawls job APIs (6 ATS platforms, the Arbeitnow board, and the Bundesagentur für Arbeit), keeps one markdown note per posting that matches your stack, and files everything else in a seen-ledger so it never resurfaces. You judge, it stamps, and prepare renders CV + cover letter (PDF/txt/md) into an upload-ready folder.

Setup — let the agent do the paperwork

The easiest way to set up amt is to not set it up yourself. Install the binaries, wire up your agent, and then just talk:

pnpm add -g @fabkho/amt        # installs the amt + amt-mcp binaries
claude mcp add --scope user amt -- amt-mcp                    # or the plugin, see below

Then tell your agent who you are — your stack, salary floor, cities or remote, what your CV should say (pointing it at an existing CV, PDF or LinkedIn export, works great as a starting point), how your cover letters should sound. It writes your profile.yaml and cv-data.<lang>.yaml into AMT_HOME (default ~/.config/amt/), which is exactly what the MCP server reads. No forms, no typing YAML by hand.

Where things land is yours to choose, via paths in profile.yaml: notesDir is where the job notes go — any folder works, and an Obsidian vault works great (the frontmatter shows up as properties, and the self-updating _index.md links every note). outputBase is where prepare drops the upload-ready application folders.

"Set up amt for me: I'm a senior React/Go dev in Hamburg, remote-only, floor 80k. Here's my CV: …"

Prefer doing it by hand? amt doctor checks the office is open (and installs Chromium for PDFs), amt init walks you through it interactively.

Der Dienstweg (the loop)

The full experience is one word to your agent: "update" — it crawls every source including the agent channels (LinkedIn & Co.), ranks everything new against your profile, and files the day's arrivals in inbox/<date>.md next to your notes (the index links every day that still has unranked mail). The same loop by hand:

amt sources add shopware   # track a company — its ATS is discovered automatically
amt crawl                  # fetch APIs + tool-crawled channels (LinkedIn, VueJobs)
amt list --status new      # today's stack of files
amt show <slug>            # one posting, full description
amt status <slug> shortlist --score 85   # der Stempel
amt prepare <slug> --lang en             # CV + letter → upload-ready folder
amt status <slug> applied                # case closed. next window, please

Statuses: new → shortlist → applied → interview (or rejected / cut — with --reason, because an Amt files even its rejections properly). Also on duty: amt import <url> for pasted ATS links (--company/--title for LinkedIn & co.). The _index.md overview in your notes directory keeps itself current.

It learns from you. Every cut carries a reason and an optional note, every score an assessment — all in the markdown, so any future session reads back why you passed on something. amt suggest (MCP: suggest_profile_updates) turns that history into proposed profile.yaml updates — companies you keep cutting, your reason mix — for you to confirm. Your blocklists and rules are the memory that hardens over time: the agent proposes, you approve.

Die Ausfertigung (CV & cover letters)

amt doesn't just find jobs — it produces the application. amt prepare <slug> builds an upload-ready folder: your CV rendered from cv-data.<lang>.yaml (one file per language, de/en) and a cover letter drafted the civilized way: the first run scaffolds cover-letter.<lang>.md, you write it with your agent in chat — carrying your tone rules from profile.yaml (the write-application prompt does exactly this) — and the second run renders everything to PDF, html, and txt. Templates are plain Nunjucks + CSS (templates/), PDFs come out of real Chromium, so what you preview is what HR gets.

Der Schalter (web dashboard)

Prefer a screen to a terminal? amt serve opens a local dashboard at http://localhost:4400 — triage the day's arrivals like a stack of files.

• 📥 Inbox — the day's new postings, ranked; accept → shortlist or reject (with a reason and a free-text note) right from the row. Postings below your score threshold are pruned automatically; rows leave the inbox as you clear them.
• Aktenboard — every note, filterable by status, remote/hybrid, city, minimum score, or free text — filters kept in the URL. Defaults to active (hides rejected/cut). Rows carry the source platform's icon; applied roles show when you applied and a button to open their documents folder.
• One original-posting click — each row opens the real ad in a new tab.

It's server-rendered (htmx, no build step, no framework), and every click writes straight to your markdown notes — so Obsidian and the dashboard never disagree. The screenshot is generated from a fictional demo dataset by pnpm shoot (scripts/shoot.mjs); design notes in docs/webview-plan.md.

Die Quellenlage (sources)

amt crawls official job APIs directly: boards like Arbeitnow, the Bundesagentur für Arbeit Jobsuche (Germany's largest job database — queried along your stack keywords and cities), and the career pages of companies you track — amt sources add <company> auto-discovers which ATS they use (Recruitee, Ashby, Greenhouse, Lever, Personio, SmartRecruiters).

Sites without an API go into sources.yaml as a channel: a recipe in the same file. A channel with a crawl: block (URL template + a parse spec — CSS selectors, regex, or JSON paths) is fetched by the tool itself, right alongside the boards; amt init seeds tool-crawled recipes for LinkedIn and VueJobs. A channel with only a free-form recipe: and no crawl: block stays agent-executed — the tool stores it and your agent runs it during a round (good for detail pages behind a bot wall, like StepStone). Add render: true to a crawl: block and amt fetches the page through the bundled Chromium for JS-rendered sites.

So amt crawl genuinely fetches everything with a machine spec — the only step that still needs the agent is judgment (ranking). Recipes are just data: tweak a selector or keyword and the very next crawl uses it, no release required.

Der Sachbearbeiter (agent integration)

The MCP server is the recommended way to use amt — the CLI is the paper form, the agent is the clerk who fills it in. amt-mcp exposes the tools (crawl, import, status, prepare, channels), job:// resources, and two workflow prompts (daily-update, write-application) carrying your profile and tone rules. The agent drafts with you in chat; the logic stays in code.

One verb, four levels of formality — they all reach the same loop:

| You say | Surface | What runs | | --- | --- | --- | | "update" | skill trigger phrase | the daily-update workflow | | /amt:daily-update | MCP prompt | crawl_jobs → channels → rank → inbox delta | | — | crawl_jobs tool | fetches APIs + tool-crawled channels, returns the rest to do | | amt crawl | CLI | the deterministic sub-step only (ranking still needs the agent) |

amt crawl deliberately isn't called update: the bare binary can't run channels or judge fit, so it fetches what it can and says what's left.

• Claude Code: /plugin marketplace add fabkho/amt → /plugin install amt@amt (bundles the MCP server + workflow skills, with auto-updates)
• Codex CLI: codex mcp add amt -- amt-mcp + copy skills/ to ~/.agents/skills/
• Cursor / Zed: point a stdio MCP server at amt-mcp
• pi: copy pi-extension/ to ~/.pi/agent/extensions/amt/

Sprechzeiten (troubleshooting)

• Node ≥ 22 and pnpm (corepack enable; if a global install complains about a missing global bin dir, run pnpm setup and restart your shell).
• amt behaves strangely / prints nonsense: macOS ships a deprecated /usr/sbin/amt. If it shadows the real one, add to the END of your ~/.zshrc: export PATH="$PNPM_HOME:$PATH" — which amt should point into your pnpm dir.

Development

pnpm install && pnpm lint && pnpm build && pnpm typecheck && pnpm test — rules and architecture in CONTRIBUTING.md.