indeed-axi

Agent-ergonomic CLI for the Indeed employer dashboard. AXI-compliant: TOON output, structured errors, help[] next steps, and a persistent authenticated browser an agent can actually drive.

Indeed exposes no public employer API, so indeed-axi is browser-only: a visible Chrome on a dedicated persistent profile (manual login, 2FA included) that stays alive between commands, plus ref-based interactive primitives modeled on playwright-axi, plus a discover command that records the dashboard's own JSON traffic for building domain commands (candidates, messages, pipeline moves) later.

Install

npm install -g indeed-axi
# or run on demand, identical:
npx -y indeed-axi@latest jobs list
# install the agent skill for coding agents (Claude Code, Codex, OpenCode, ...):
npx -y indeed-axi skill install

Requires Node 20+ and a Chrome/Chromium. By default the daemon runs the Playwright-installed browser when available (on macOS its own "Google Chrome for Testing.app" - real Chrome engine in a separate app bundle, fully isolated from your daily Chrome; install with npx playwright-core install chromium), falling back to installed Google Chrome. Override with INDEED_BROWSER (auto default, chromium, or chrome) or INDEED_BROWSER_BIN (explicit binary path; always wins).

Auth

There is nothing to configure - no API key exists. Log in manually once:

indeed-axi auth login    # opens visible Chrome; you log in (2FA included)
indeed-axi auth status --browser

The session lives in a dedicated persistent profile under ~/.indeed-axi/ (INDEED_STATE_DIR overrides) - never your primary Chrome profile. The browser stays open after login so browser commands work immediately.

Commands

| Command | Description | | --- | --- | | indeed-axi | Home: session record, daemon state, next steps (zero launches) | | indeed-axi jobs list | Live employer jobs (short ref + title) | | indeed-axi candidates list --job <ref> | Live pipeline rows (--stage new\|reviewed\|...) | | indeed-axi candidates sync --job <ref> | Pull new/changed candidates into the local store (--applications) | | indeed-axi candidate get <id> | One candidate packet (store, or live with --refresh) | | indeed-axi messages read <id> | Full candidate thread, out/in roles (--full for bodies) | | indeed-axi messages send <id> --text "..." | Preview a single message send (--confirm to actually send, then verify by thread re-read) | | indeed-axi messages unread | Live unread conversation count | | indeed-axi stage move <id> --to <milestone> | Preview a pipeline move (--confirm to apply; new…rejected, hired) | | indeed-axi candidates note <id> --text "..." | Preview an employer-internal note (--confirm to create) | | indeed-axi digest | Ranked hiring review over the local store (offline): scores, awaiting-reply aging with chase/expire flags | | indeed-axi score record <id> --stage --score --rationale | Record a screening score locally; --note --confirm syncs it into Indeed | | indeed-axi auth login | Open visible Chrome, wait for manual login | | indeed-axi auth status | Session state (--browser probes live) | | indeed-axi auth logout | Close the browser, clear the record (--purge deletes the profile) | | indeed-axi browser open | Start/attach the browser, navigate, snapshot | | indeed-axi browser status / close | Daemon liveness / stop the browser | | indeed-axi browser goto <url> | Navigate to an indeed.com URL (scoped) | | indeed-axi browser snapshot | Page outline with [ref=eN] tags (--full, --query) | | indeed-axi browser find <terms> | Filtered outline lines with context | | indeed-axi browser click <ref> | Click a ref, get a fresh snapshot | | indeed-axi browser fill <ref> <text> | Fill a field (--submit presses Enter) | | indeed-axi browser select <ref> <value> | Choose a dropdown option | | indeed-axi browser press <key> | Press a key (Enter, Tab, Control+A, ...) | | indeed-axi browser eval <expr> | Evaluate JS on the page, JSON result | | indeed-axi browser screenshot | Save a PNG (runs dir by default) | | indeed-axi browser console | Collected in-page errors | | indeed-axi discover | Record a manual dashboard session (dev command) | | indeed-axi setup hooks | Ambient session context (Claude Code, Codex, OpenCode) |

Every command supports --json and --help.

Domain reads (GraphQL through the session)

Domain commands run over the same authenticated browser: at connect time the CLI sniffs the exact headers the dashboard itself sends to apis.indeed.com/graphql (in memory only), then performs in-page fetches. Candidate packets (submission, application answers, thread refs) persist to ~/.indeed-axi/store/candidates/<id>.json — local PII state, never committed. Syncs are diff-based and paced (~300 ms between per-candidate fetches).

$ indeed-axi jobs list
jobs:
  count: 3
list[3]{ref,title}:
  bbb2c3d4-5e6f-4701-928a-23456789abcd,Content Producer & Social Media Manager
  ...

$ indeed-axi candidates sync --job "content producer" --applications
sync:
  job: Content Producer & Social Media Manager
  found: 3
  fresh: 3
  fetched: 3
list[3]{id,name,milestone}:
  aaaaaaaa0003,Sam Ortega,REVIEWED
  ...

$ indeed-axi messages read aaaaaaaa0002
messages:
  count: 2
list[2]{role,at,body}:
  out,"2026-09-29T15:01:16Z","Hi Jordan, Thanks for submitting your..."
  in,"2026-09-29T15:16:35Z","Hi Bryce! Thank you for getting back to me!..."

The interactive loop

$ indeed-axi browser open
browser:
  started: true
  port: 54525
page:
  url: "https://employers.indeed.com/c/dashboard"
snapshot:
  lines: 89
  refs: 13
outline[89]:
  - link "Candidates" [ref=e1]
  - button "Message" [ref=e2]
  ...

$ indeed-axi browser find candidate     # narrow the outline
$ indeed-axi browser click e1           # act, fresh snapshot comes back
$ indeed-axi browser fill e7 "..." --submit

Refs come from the most recent snapshot and reset on every snapshot or navigation - stale refs fail loudly as STALE_REF. Elements that are not currently in the viewport appear as (offscreen) without a ref: scroll (browser press End, eval window.scrollBy(0, 800)) and re-snapshot to make them actionable. Ghost elements (hidden nav clones) never get refs.

Browser daemon

All commands share one headed Chrome (the daemon): launched detached on the persistent profile with a local DevTools port, reattached over CDP per command. browser close (or auth logout) stops it; a stale state file is self-healing. One browser at a time, always headed, always Indeed-only:

• Navigation and actions are scoped to *.indeed.com.
• Cookies never leave the browser; nothing is exported or replayed.
• Captured discovery artifacts pass through redaction (no headers, cookies, or credential-shaped keys).

Discovery

discover records a manual walkthrough of the employer dashboard (open a candidate list, open a candidate, read a thread, send a message, move a stage) and writes trace.zip (when available), network.json, documents/ (captured JSON bodies), console-errors.txt, final screenshot, aria.yml, dom.html, and index.json under ~/.indeed-axi/runs/discover-<id>/. Those artifacts feed the domain-command roadmap (candidates, messages, pipeline moves with --confirm gates).

Gated writes (Phase 2)

Every mutation is a plan-preview by default: the command resolves the candidate live, shows exactly what would happen, and makes zero mutation calls. Rerunning with --confirm performs the write and verifies it (sends by an independent thread re-read, stage moves against the mutation's own milestone response). One send per invocation - never bulk.

$ indeed-axi messages send aaaaaaaa0002 --text "Quick follow-up: ..."
send:
  candidate: {id: aaaaaaaa0002, name: Jordan Solis, milestone: REVIEWED}
  chars: 22
message: "Quick follow-up: ..."
help: Preview only - zero messages were sent,Rerun with --confirm to send this message

$ indeed-axi stage move aaaaaaaa0003 --to phone_screened --confirm
moved:
  to: PHONE_SCREENED
  verified: true

Logged-out sessions fail fast (~5s) with an auth login hint instead of timing out on traffic capture.

Roadmap

Full requirements live in REQUIREMENTS.md.

• Phase 0 - scaffold, persistent browser daemon, auth, interactive browser family, discover, skill, hooks.
• Phase 1 - domain reads over in-page GraphQL: jobs list, candidates list/sync, candidate get, messages read/unread, local candidate store. Endpoint map: docs/discovery-2026-09-29.md.
• Phase 2 - gated writes: messages send --confirm (verified by thread re-read), stage move --confirm, candidates note --confirm (employer-internal notes - the score-sync channel).
• Phase 3 (this release) - workflow layer: score record (local + --note --confirm sync), digest (offline ranked review with chase/expire aging), thread-state caching (sync --threads, messages read), plain-text application extraction, sync --force.
• Future - session keepalive / automated login: Indeed sessions are short-lived and auth login is manual today. Investigate a safe credential-handling design (e.g., a keychain-backed login flow) - manual login remains the doctrine until then.

Set up on another machine

npm install -g indeed-axi
npx -y indeed-axi skill install      # writes the agent skill to ~/.agents/skills/
indeed-axi auth login                # visible Chrome; log in manually (2FA included)

The skill installer is idempotent (--force overwrites a changed version, --dir <path> targets another skills directory). Optional env: INDEED_STATE_DIR (default ~/.indeed-axi), INDEED_BROWSER (auto default, chromium, or chrome), and INDEED_BROWSER_BIN (explicit binary path; always wins).

Candidate packets live under ~/.indeed-axi/store/ per machine and are never synced - sync them live with candidates sync instead. Indeed sessions are short-lived; when data commands report AUTH_ERROR, rerun auth login.

Development

npm install
npm run build        # tsc -> dist/ (committed; rebuild on every change)
npm test             # vitest, fully mocked - no browser needed
npm run typecheck

Run locally with node bin/indeed-axi.js. Network/browser access is isolated to src/browser/; commands receive injected daemon connections and tests pass fakes. See AGENTS.md for the full house rules.

License

MIT