boardwatch

A self-hosted job radar that reads the fine print. Point it at the companies you care about. boardwatch watches their official ATS job boards, catches new postings early, and ranks them against your profile with an explainable score. It also reads each posting for the hard eligibility requirements that quietly rule people out (visa sponsorship, security clearance, a required degree, years of experience, a location) and flags the ones you could not actually apply to, each backed by the exact sentence it read as evidence. Nothing is guessed, nothing phones home, and it all runs on your own machine.

Status: pre-release, under active development. boardwatch does not submit applications: it finds and evaluates postings, and you decide what to do with them. No telemetry. No accounts, no API keys. Your data stays in a local SQLite file.

$ boardwatch top
 #   Title                            Company     Score   Eligible   Why
 12  Senior Backend Engineer          Stripe      0.86    no flags   covers 9/11 skills · title · 1d
 7   Software Engineer, Platform      Linear      0.81    check      covers 7/10 skills · title · 3d
 33  Backend Engineer (Payments)      Ramp        0.74    no flags   covers 6/9 skills · 2d
 5   Full-Stack Engineer              Supabase    0.68    no flags   covers 5/8 skills · title · 6d
 18  Infrastructure Engineer          OpenAI      0.61    check      covers 4/9 skills · 4d

2 postings hidden as ineligible (run with --include-ineligible to see them).

(Illustrative output. # is the posting id; pass it to boardwatch show <id> for the full posting, a per-component score breakdown, and the eligibility audit with quotes. "no flags" means no catalogued disqualifier was found, not that you are cleared to apply; "check" means the posting was ambiguous.)

Why boardwatch?

Job boards optimize for their advertisers, not for you. LinkedIn/Indeed bury fresh roles under sponsored noise and stale reposts; paid trackers put a subscription (and their servers, and your search history) between you and postings that are already public.

boardwatch takes the direct route. Greenhouse, Lever, Ashby, Workable, and SmartRecruiters each expose a public, keyless JSON endpoint for every board they host, the same data the company's own careers page renders. boardwatch polls those endpoints politely, on your schedule, and tells you what's new since last time.

Honest limits. boardwatch only covers companies hosted on Greenhouse, Lever, Ashby, Workable, or SmartRecruiters (a large slice of tech, but not everyone, no Workday/Taleo/etc. yet). It reads exactly what those APIs expose. It is pre-release: expect rough edges, and read Responsible use before pointing it at boards you don't own.

Quickstart (≈2 minutes to your first shortlist)

pipx (recommended)

pipx install boardwatch      # isolated, on your PATH
boardwatch init              # pick a starter set of companies + paste your profile
boardwatch scan              # poll the watched boards (polite, conditional GETs)
boardwatch top               # ranked shortlist

boardwatch init is interactive. Pick [1] Starter set to watch a curated group of well-known boards in one keystroke, [2] Search registry to pick from the bundled catalog, or [3] Paste any provider:slug or board URL. Then paste your résumé text and a few target/exclude titles and locations. A cold init → scan → top lands your first ranked shortlist in well under ten minutes.

Docker

docker run --rm -v boardwatch-data:/data \
  ghcr.io/mit112/boardwatch:latest --data-dir /data init
docker run --rm -v boardwatch-data:/data \
  ghcr.io/mit112/boardwatch:latest --data-dir /data scan
docker run --rm -v boardwatch-data:/data \
  ghcr.io/mit112/boardwatch:latest --data-dir /data top

From source

git clone https://github.com/mit112/boardwatch && cd boardwatch
uv sync                      # https://docs.astral.sh/uv/
uv run boardwatch init

How it works

   init ──▶ scan ──▶ top ──▶ show <id>
   │        │        │        │
   companies fetch    rank on   full posting +
   + profile boards   demand    score breakdown

• init: one-time setup. Choose companies (starter set / registry search / paste), then your profile (résumé text, target titles, excludes, locations, remote-only).
• scan: fetches every watched board through a polite fetcher (per-host pacing, retries with backoff, conditional If-None-Match/If-Modified-Since so unchanged boards cost a 304), then applies each board transactionally. Prints a one-line summary.
• top [N]: ranks open postings against your profile right now (weights are read live), newest-and-most-relevant first, with a one-line "why".
• show <id>: the full posting plus a per-component score table (skill coverage, title match, recency, location fit) and the eligibility audit with quoted evidence.
• eligibility: facts / policy to tell boardwatch your situation, run to evaluate open postings, summary for a funnel of what the catalog matched. See Eligibility audit.
• companies: add / remove / search / list / import / export your watched boards. boardwatch companies add https://boards.greenhouse.io/acme just works.
• doctor: per-board connectivity and freshness, plus a local DB integrity check.
• config show / config set: tune politeness and ranking weights (see below).

Ranking, briefly

Each posting gets a 0–1 score: a weighted blend of skill coverage, title match (fuzzy), recency (exponential decay), and location fit. Undefined components renormalize away, so a sparse profile still ranks sensibly. Nothing is precomputed: change a weight and the next top reflects it. show <id> prints the exact arithmetic.

Eligibility audit

Ranking tells you how well a posting fits. The eligibility audit tells you whether you could apply at all, and shows its work.

You describe your situation once, in the catalog's own vocabulary:

boardwatch eligibility facts set work_authorization.status citizen
boardwatch eligibility facts set highest_degree bachelor
boardwatch eligibility policy set visa_sponsorship blocker   # treat this family as disqualifying

Then boardwatch eligibility run reads each open posting for catalogued requirements (visa sponsorship, security clearance, degree, years of experience, location, and more), resolves them against your facts, and stores a verdict. boardwatch show <id> prints it with the receipts:

Eligibility: ineligible
  unmet · required: work authorization / sponsorship
      quote: "This role is not able to sponsor employment visas now or in the future."

Three properties are deliberate:

• Evidence-linked. Every requirement carries the exact sentence it was read from, sliced from the posting version that was evaluated. Nothing is paraphrased or invented.
• Deterministic. The same posting and the same facts always produce the same verdict. There is no language model in the loop and nothing to hallucinate: the vocabulary and rules are a versioned catalog, and a verdict is invalidated and recomputed only when your profile or the catalog changes.
• Honest. A clean posting reads as "no flags", never as a guarantee. "no flags" means only that no catalogued disqualifier was found, not that you are cleared to apply, and ambiguous wording reads as "check" rather than a false all-clear.

boardwatch eligibility summary shows the funnel (how many postings were evaluated, the verdict split, and what fired per family) so you can watch the catalog working before you trust a hidden count. By default top hides postings that are ruled out and reports the count; top --include-ineligible shows them.

What changed since you last looked

boardwatch scan records every appearance, disappearance and body revision in an append-only ledger. digest reads the ledger from wherever you left off:

boardwatch digest          # new, reopened, updated, and a closed count
boardwatch digest --peek   # the same view without consuming it
boardwatch top --new       # rank only the postings first seen since your last digest

The cursor is an event id, not a timestamp, so a clock change or a missed day cannot skip or repeat a window.

Your funnel

boardwatch never applies for you. It records what you did, so the state stays yours:

boardwatch track add 42                    # start tracking a posting
boardwatch track status 1 applied          # move it, with an immutable ledger entry
boardwatch track status 1 interviewing --note "phone screen booked"
boardwatch track list --status applied
boardwatch track log 1                     # the full history for one application

Take your data with you

boardwatch export --format jsonl --out postings.jsonl
boardwatch export --format csv

Every row carries the posting, your funnel state, and the eligibility verdict together with the profile and rules hashes that identify the evaluation it was computed under. This is a flat snapshot, not a full audit trail; it does not support independent recomputation of verdicts.

Configuration

boardwatch config show prints every key, its value, and its default; boardwatch config set <key> <value> changes it (validated at set time and load time).

See docs/configuration.md for details.

Schedule scans

Run boardwatch init once interactively before scheduling scans. The scheduler must run as the same user that ran init, so it reads the same local profile and database. Start with a daily scan; the default politeness settings are designed for that cadence.

Each example below appends the scan summary to a log. Replace /absolute/path/to/boardwatch with the output of command -v boardwatch, then run the command once manually before enabling its timer.

cron (Linux or macOS)

Create a log directory, then add a daily job with crontab -e:

$ mkdir -p "$HOME/.local/state/boardwatch"

# Run every day at 08:00 local time.
0 8 * * * /absolute/path/to/boardwatch scan >> "$HOME/.local/state/boardwatch/scan.log" 2>&1

Cron has a deliberately small environment. If you set BOARDWATCH_DATA_DIR or BOARDWATCH_CONFIG_DIR when running boardwatch normally, define the same values above the job in the crontab.

launchd (macOS)

Save this as ~/Library/LaunchAgents/com.boardwatch.scan.plist, replacing the boardwatch path and the home-directory placeholder. The standard output and error paths must use absolute paths.

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
  <dict>
    <key>Label</key>
    <string>com.boardwatch.scan</string>
    <key>ProgramArguments</key>
    <array>
      <string>/absolute/path/to/boardwatch</string>
      <string>scan</string>
    </array>
    <key>StartCalendarInterval</key>
    <dict>
      <key>Hour</key><integer>8</integer>
      <key>Minute</key><integer>0</integer>
    </dict>
    <key>StandardOutPath</key>
    <string><home>/Library/Logs/boardwatch-scan.log</string>
    <key>StandardErrorPath</key>
    <string><home>/Library/Logs/boardwatch-scan.log</string>
  </dict>
</plist>

Load it and confirm its status:

$ launchctl bootstrap "gui/$(id -u)" ~/Library/LaunchAgents/com.boardwatch.scan.plist
$ launchctl print "gui/$(id -u)/com.boardwatch.scan"

After editing the file, reload it with launchctl bootout "gui/$(id -u)/com.boardwatch.scan" followed by the bootstrap command above.

systemd user timer (Linux)

Create ~/.config/systemd/user/boardwatch-scan.service:

[Unit]
Description=Scan watched job boards with boardwatch

[Service]
Type=oneshot
ExecStart=/absolute/path/to/boardwatch scan

Then create ~/.config/systemd/user/boardwatch-scan.timer:

[Unit]
Description=Run boardwatch scan every day

[Timer]
OnCalendar=*-*-* 08:00:00
Persistent=true

[Install]
WantedBy=timers.target

Enable the timer and inspect the most recent run:

$ systemctl --user daemon-reload
$ systemctl --user enable --now boardwatch-scan.timer
$ systemctl --user list-timers boardwatch-scan.timer
$ journalctl --user -u boardwatch-scan.service --since today

Persistent=true runs a missed daily scan after the next login. To keep user timers running after logout, enable lingering for the account with loginctl enable-linger "$USER".

Notifications

notify is a standalone command, a sibling of scan: chain them in your scheduled job so each run scans, then pushes anything new:

/absolute/path/to/boardwatch scan && /absolute/path/to/boardwatch notify

Notifications are off by default. Turn on one or both channels:

boardwatch config set notify.webhook_enabled true
boardwatch config set notify.desktop_enabled true

The webhook channel needs a URL from the environment, never from config.toml:

export BOARDWATCH_NOTIFY_WEBHOOK_URL=https://hooks.slack.com/services/...

One payload works for Slack incoming webhooks, Discord webhooks, and generic/structured consumers, so the same URL drops into any of them. Desktop notifications are best-effort (macOS via osascript, Linux via notify-send); on any other platform, or if the notifier binary is missing, desktop delivery degrades non-fatally and webhook remains the cross-platform, headless-friendly channel. Run boardwatch notify --dry-run to preview what would be sent without delivering anything or advancing the notify cursor.

Tailor a résumé

boardwatch never parses a résumé — parsing free-form PDFs/Word docs to understand a person's history is exactly the kind of guessing this project avoids. Instead you author your résumé once as structured YAML, and tailor only ever renders it:

boardwatch tailor init                 # scaffold {config_dir}/resume.yaml, edit it in your editor
boardwatch tailor validate             # confirm it loads; see entry/bullet counts + detected skills
boardwatch tailor run <posting-id>     # tailor it against one posting's extracted JD skills

validate and run read {config_dir}/resume.yaml unless you pass --resume PATH. run also takes --out DIR (default {data_dir}/tailored), --format typst (the only 1.0 adapter), and --dry-run (report only; writes no file and records no artifact). It prints one line per bullet — kept, reordered, swapped, or dropped, with the JD skills that bullet covers — and the same per-bullet audit is stored on the artifact row.

tailor run selects which of your authored bullets to keep (favoring ones that cover the posting's extracted skills), reorders skill mentions, and applies whole-token synonym swaps from a small, bundled, frozen equivalence table (e.g. "JS" ↔ "JavaScript") — never free text generation. Before anything is written, a no-fabrication guarantee re-checks the output against your master résumé: every kept bullet's tokens must be either unchanged or a substitution the equivalence table names, and no bullet, entry, or section can appear that wasn't in the original. A résumé that fails this check is rejected before any file or database row is written, never delivered as a "best effort".

Honest bounds (Tier A). This is Tier A: a local, deterministic bullet-selection and safe-synonym pass — it does not rewrite your prose, invent new claims, or call any model. Your profile text (the free-form blurb from boardwatch init) is never imported into the résumé — tailor reads only what you author in resume.yaml. PDF output is best-effort: it shells out to a local Typst install if present; without one you still get the rendered Typst source and can compile it yourself, or paste it elsewhere. Output lands at {data_dir}/tailored/tailored-<posting-id>.{typ,pdf} — a deterministic path, so re-running tailor run for the same posting overwrites that file even though each run is recorded as its own artifact in the database; the file on disk always reflects your most recent run, not necessarily the one you're currently reading about. If a later Typst compile fails, the stale PDF from the previous run is removed rather than left behind next to the new source.

Tier B (opt-in LLM)

Tier A never rewrites your prose. If you want bullets reworded toward a posting's language, Tier B is an opt-in, off-by-default LLM lane on top of it:

boardwatch tailor run <posting-id> --tier-b     # alias: --llm

--tier-b requires all of the following, and does nothing (writes nothing, exits 1) if any is missing:

• llm.resume_tailoring = true and llm.enabled = true in {config_dir}/config.toml — resume_tailoring is the only key Tier B adds, and it lives on the same [llm] block as the opt-in LLM eligibility-extraction tier (see docs/configuration.md); no new section, no new secret. config set llm.* is reserved and refuses to write these, so set them by editing config.toml directly.
• BOARDWATCH_LLM_API_KEY in the environment (never in config.toml).

Per bullet, Tier B proposes a reworded version, then runs it through a deterministic overmatch filter and a fail-closed entailment judge (the judge sees only the two bullet texts — original and reworded — and never the job description). A bullet is only kept reworded if it passes both; otherwise Tier B silently falls back to the Tier A text for that bullet, so a --tier-b run degrades to Tier A on any single bullet without failing the whole command. The CLI reports how many bullets were reworded vs. fell back, and why.

Tier B costs 2 LLM calls per bullet (propose, then judge), drawn from the same llm.max_calls_per_run budget (default 50, shared with the eligibility LLM lane) — so roughly 25 bullets per run before the tail starts falling back with drop_reason: "budget". That budget is consumed even on a cache hit, by design, so re-running the same posting does not extend it; raise llm.max_calls_per_run in config.toml instead. See docs/configuration.md.

Dual output, not a replacement. --tier-b always writes the ordinary Tier A file first — Tier B never runs in place of it — plus a second file/artifact (resume_tailored_llm) with reworded bullets marked // reworded (Tier B) in the rendered source. The lineage is recorded as B —rewritten_from→ A —tailored_from→ your master résumé, so either output can be traced back.

Honest bounds (Tier B). Tier B is not the no-fabrication guarantee above: passing the filter and judge is evidence, not proof, and every reworded bullet is meant to be read by you before you send it, not trusted blind. The guarantee is layered: the deterministic filter rejects invented skills, added/inflated numbers, and invented brand/company names outright, provider-independently; the entailment judge is a best-effort, model-dependent second layer for the fabrications a text-diff can't see (flipped negations, inflated seniority, swapped outcomes); and you are the backstop. The Tier A path never calls out to a model, regardless of Tier B's settings; only --tier-b sends bullet text and the posting's extracted JD skill names to the configured provider, and only when explicitly enabled and requested. See SECURITY.md for exactly what leaves your machine and when.

Tier B without an API key (agent lane)

--tier-b needs BOARDWATCH_LLM_API_KEY and a metered provider. If you only have a Claude Code subscription and no API key, there's a second, subscription-driven Tier B lane that gets the same reworded-bullet output by having Claude Code itself do the rewrite, with boardwatch validating it through the identical filter + judge + Tier A stack described above.

It's a three-command handshake, driven by the tailor-rewrite boardwatch skill (.claude/skills/tailor-rewrite):

boardwatch tailor rewrite request <posting-id>                                     # 1. writes rewrite_request.json
# the skill's rewriter agent reads it and writes candidates.json
boardwatch tailor rewrite screen <posting-id> --candidates candidates.json         # 2. writes judge_request.json
# a SEPARATE, freshly scoped judge subagent reads it and writes verdicts.json
boardwatch tailor rewrite apply <posting-id> --candidates candidates.json --verdicts verdicts.json  # 3.

request runs Tier A internally and hands the rewriter agent each bullet plus the posting's JD skills. screen re-derives the authoritative bullet text from a fresh Tier A run and puts the rewriter's candidates through the same deterministic overmatch filter as the API lane — but the judge_request.json it writes is JD-blind by construction: it's built from only (original, candidate) pairs, with no job description or skills field anywhere in it, and boardwatch's skill instructs the judge to run as a separate subagent so it never inherits the rewriter's JD-aware context. apply parses the judge's verdicts with the same exact-token allowlist as the API lane and emits both artifacts.

Gate: llm.resume_tailoring_via_agent = true in {config_dir}/config.toml. Unlike --tier-b, this lane needs no llm.enabled and no API key — boardwatch makes no LLM call itself in this lane; the rewriting and judging happen in Claude Code subagents outside the CLI. (config set llm.* is reserved, same as above — hand-edit config.toml.)

The per-bullet call budget is advisory here, not a hard spend limit: subscription calls aren't API-metered, so it's set wide enough to never truncate a legitimate run and functions only as a soft cap on how many bullets get attempted in one pass. Every other safety backstop is unchanged: the Tier A file is always emitted alongside, every reworded bullet is marked // reworded (Tier B), and the artifact is meant to be reviewed by you before it's sent — this lane is LLM-assisted and filter+judge gated, not structurally proven, same as --tier-b. Provenance on the artifact records provider claude-code-agent, model subscription, so it's distinguishable from an API-lane run.

Supported boards

boardwatch ships a bundled registry of verified public boards (35+ companies, with a curated starter set), so init works offline out of the box. You can watch any board these providers host, not just the registry, with companies add. The registry is community-maintainable by PR; see src/boardwatch/registry/README.md.

SmartRecruiters honest limits. Its API cannot distinguish a typo'd company slug from a real, empty board — an unknown company returns an empty board, not an error, so companies add and doctor flag it as unverifiable rather than confirmed. Job bodies are fetched once per posting (bounded by detail_fetch_budget, default 50) and never refreshed, since the list endpoint carries no revision signal for description-only edits. A posting that goes inactive while still listed is not re-detected as closed until it drops off the list — it self-heals on a later scan.

Responsible use & legality

boardwatch reads the same public, keyless endpoints that power each company's own careers page: it does not scrape rendered HTML, log in, or bypass any access control. That is deliberately the least-invasive way to get this data. Still, these are third-party services, and using them responsibly is on you:

• Keep the politeness defaults. The defaults (≥1 request/sec per host, conditional GETs, bounded retries, a descriptive User-Agent) are intentionally gentle. Don't crank scan_workers up or per_host_delay_seconds down to hammer a board.
• It's for personal job-search use, not bulk data resale or redistribution of posting content. boardwatch stores postings locally for your review.
• Provider terms & rate limits can change and may restrict automated access. You are responsible for complying with each provider's Terms of Service. If a provider asks you to stop, stop.
• No warranty. These are undocumented-stability public endpoints; they can change or break without notice.

If you're unsure whether your use is appropriate, err toward watching fewer boards, less often. A job seeker checking a dozen companies once a day is the intended shape.

Privacy & data

• Local-first. Its primary store is one SQLite database in your platform data directory; the opt-in LLM tier also caches raw responses there as plain files on disk (override with --data-dir). No server, no account, no cloud.
• No telemetry. boardwatch phones home to nobody.
• Two optional secrets, env-only. The default path authenticates to nothing. The opt-in LLM tier reads BOARDWATCH_LLM_API_KEY and the opt-in webhook notifier reads BOARDWATCH_NOTIFY_WEBHOOK_URL, both from the environment only and never written to disk. See SECURITY.md.

Roadmap

• PyPI + GHCR published releases (pipx install boardwatch, docker run …)
• Notifications on new matches (desktop / webhook)
• digest and top --new change detection (only what changed since last run)
• More ATS providers (community-driven)
• Data-portability export (--format jsonl|csv)
• Résumé tailoring (tailor init/validate/run, local, no-fabrication guarantee)

Next:

• Workday provider (host/tenant/site boards, the largest ATS still out of reach)
• More eligibility rule families (contract vs. full-time, internships)
• A readable settings surface, so every opt-in feature is discoverable and reversible without hand-editing config.toml

These boxes track main, which can run ahead of the latest published release. CHANGELOG.md is the authoritative record of what shipped in the version pipx install boardwatch gives you; anything under Unreleased needs an install from source.

Have a company on a board boardwatch doesn't reach yet, or an ATS you want supported? Open an issue.

Contributing

Contributions welcome: code, registry entries, or bug reports. See CONTRIBUTING.md for dev setup (uv sync, make check) and the registry guide for adding a company board. All changes land via PR against a branch-protected main.

License

MIT.