Witness

Witness does not auto-apply to jobs. It never submits anything for you — there is no submit button in the codebase. It does the opposite of a spam bot: it makes each application fewer, better, and unfakeable, then hands it to you to send.

An AI job tool that refuses to lie for you — and refuses to hit submit for you.

It pulls a company's real application questions (from public Greenhouse/Ashby APIs, not a guessed form), traces every claim in your draft back to a real line in your CV (a proof-map — over-claiming is a bug it catches, not a feature it maximizes), scores drafts for genericness before you ever see them, and writes it all to an offline dashboard you click through yourself. Built for Claude Code (MCP + skills) and as a CLI.

Adversarially tested against 40 failure modes. If a company's real form can't be loaded, you see a fetch_failed flag and a red banner — Witness does not fake a form and call it done.

---

Why it's different

• Real ATS fields, not guesses. Greenhouse + Ashby return the exact questions a company asks (custom ones included). Others fall back to a standard set — and say so. Fetch failures are surfaced, never hidden.
• Every claim has a receipt. The proof-map binds each statement in a draft to a real fact in your CV; an honesty check flags anything unbacked. (AI-inflated resumes now get detected ~76% of the time and auto-rejected by half of hiring managers — Witness is built so you're never the one caught.)
• You submit, it never does. Human-in-the-loop is an enforced design contract (see TRADEMARK.md / CONTRIBUTING.md), not an MVP gap. No auto-apply, no proxy/CAPTCHA bypass, no fake accounts — ever.

Plus a non-trivial layer nothing else does: it reframes a role onto the axis where you win and drafts a short behavioural teardown of the company's own product — a witness only you could produce.

Your data folder

Witness is local and bring-your-own-data. Point it at a folder with your CV-derived profile; outputs are written back into that same folder. Nothing personal lives in this repo.

my-jobsearch/
  profile.json          # from `parse-cv`, or copy examples/profile.example.json
  config.json           # optional; copy examples/config.example.json
  output/jobs.json      # written by Witness
  applications/*.json   # written by Witness (+ a non-trivial kit each)
  dashboard.html        # written by Witness — open it or serve on localhost

Quickstart (60 seconds)

python3 -m venv .venv && source .venv/bin/activate        # Python 3.10+
pip install -r requirements.txt
export ANTHROPIC_API_KEY=…                                 # optional (LLM CV-parsing + drafted answers)

D=~/my-jobsearch
python -m src.cli --data-dir $D parse-cv --cv ~/MyCV.pdf   # CV -> profile.json
python -m src.cli --data-dir $D scrape                     # -> output/jobs.json
python -m src.cli --data-dir $D apply --url "<job_url>"    # -> applications/<company>.json (+ kit)
python -m src.cli --data-dir $D dashboard                  # -> dashboard.html
python -m http.server -d $D 8899                           # then open http://localhost:8899/dashboard.html

Claude Code (MCP + skills)

Add the server from .mcp.json and set JOBSEARCH_DATA to your folder, then just ask Claude: "find jobs", "apply to <link>", "make this application non-trivial", "build my dashboard". The MCP server (src/mcp_server.py) exposes get_profile, read_jobs, refresh_jobs, get_application_fields, generate_application, make_nontrivial, score_triviality, build_dashboard, list_companies. Skills live in .claude/skills/.

Architecture

CV ─parse_cv─▶ profile.json ─┐
job boards ─scrape_jobs(JobSpy)─▶ output/jobs.json
job URL ─ats(Greenhouse/Ashby public APIs)─▶ exact fields ─┐
                                      generate_application ─┴─▶ applications/*.json (+ nontrivial kit)
                                                build_dashboard ─▶ dashboard.html (offline, copy buttons)

Tests

python -m tests.adversarial          # offline (40 checks)
python -m tests.adversarial --net    # + live ATS probes

Using it responsibly

Job discovery uses JobSpy to read public listings. LinkedIn, Naukri, Indeed and others prohibit automated access in their Terms of Service. Witness makes no warranty of ToS compliance; whether and how you run discovery is your own decision and responsibility. A conservative delay is on by default — keep it, and don't run Witness as a scheduled/always-on service. Full terms in DISCLAIMER.md. Design invariants (no auto-submit, no circumvention) in CONTRIBUTING.md.

License

Apache-2.0. The name and its human-in-the-loop guarantee are covered by TRADEMARK.md — fork the code freely under a different name; don't ship a guardrail-stripped build as "Witness".