a-jobseeker

A command line tool that:

1. Scrapes job offers from various websites.
2. Selects the ones that match your profile, using an AI.
3. Generates a tailored PDF CV for each selected offer.
4. Outputs the list (application link + CV).

Installation

• From PyPI:

pip install a-jobseeker

On systems protecting their Python installation (Debian, Ubuntu...), install it with uv (uv tool install a-jobseeker) or pipx (pipx install a-jobseeker) instead.

• Or from the GitHub repository:

mkdir -p ~/.local/src
cd ~/.local/src

REPO_NAME=a-jobseeker

git clone https://github.com/Skalyaev/$REPO_NAME.git
cd $REPO_NAME && pip install .

An AI command line program is required. Claude Code (claude) is used by default, but any other AI CLI can be plugged in (see AI providers).

Quick start

a-jobseeker init # creates example config.json and profile.json files in ~/.config/a-jobseeker

# ... edit profile.json (your profile) and config.json (your searches)

a-jobseeker run # scrapes, selects and generates the applications

Example output:

1 offer(s) matching your profile:

[1] Software Engineer (H/F) - CDI - Paris - Wiremind (Paris)
    Score : 85/100
    Why   : Junior fullstack role in Paris matching the candidate's Python and FastAPI experience.
    Apply : https://www.linkedin.com/jobs/view/4446409330/
    CV    : ~/.local/share/a-jobseeker/applications/2026-09-18/linkedin-Wiremind-Software-Engineer-4446409330/cv.pdf

Commands

Options available on every command:

run options:

Files and directories

When set, $XDG_CONFIG_HOME, $XDG_DATA_HOME and $XDG_CACHE_HOME replace ~/.config, ~/.local/share and ~/.cache. The command line options take precedence over both.

Thanks to seen.json, a periodic run (cron) only evaluates new offers. Thanks to the cache, offers that could not be evaluated yet are not downloaded again. The cache can be deleted at any time.

The profile (profile.json)

The profile is sent as-is to the AI: its structure is free, add anything that may help.

Only these fields are read by the program itself:

• identity: copied verbatim into the CV, never generated by the AI. first_name is required, every other field is optional.

• preferences.banned_companies and preferences.banned_title_keywords: filters applied before calling the AI (whole word, case insensitive, on the offer title).

Everything else is only read by the AI, to select the offers and write the CV. The example below shows every key of a complete profile:

{
  "identity": {
    "first_name": "Jane",
    "last_name": "Doe",
    "email": "jane.doe@example.com",
    "phone": "+33 6 00 00 00 00",
    "location": "Paris, France",
    "links": [
      { "label": "Portfolio", "url": "https://jane-doe.example.com" },
      { "label": "LinkedIn", "url": "https://www.linkedin.com/in/jane-doe" },
      { "label": "GitHub", "url": "https://github.com/jane-doe" },
    ],
  },
  "headline": "Backend developer",
  "summary": "Backend developer with 3 years of experience building high-traffic Python APIs.",
  "objectives": "Join a product team working on data and performance topics.",
  "experiences": [
    {
      "title": "Backend developer",
      "company": "Example Inc.",
      "company_description": "Logistics SaaS vendor",
      "contract": "Permanent",
      "location": "Paris",
      "start": "2023-01",
      "end": "present",
      "description": "Design and maintenance of the parcel tracking APIs.",
      "achievements": [
        "Cut the tracking API response time by 40% with a Redis cache.",
        "Set up CI/CD with GitLab CI and Docker.",
      ],
      "technologies": [
        "Python",
        "FastAPI",
        "PostgreSQL",
        "Redis",
        "Docker",
        "GitLab CI",
      ],
    },
  ],
  "projects": [
    {
      "name": "sync-tool",
      "date": "2024-06",
      "description": "Command line file synchronization tool.",
      "technologies": ["Go"],
      "url": "https://github.com/jane-doe/sync-tool",
    },
  ],
  "skills": [
    { "category": "Languages", "items": ["Python", "Go", "SQL"] },
    { "category": "Tools", "items": ["Docker", "PostgreSQL", "Redis", "Git"] },
  ],
  "education": [
    {
      "degree": "MSc Computer Science",
      "school": "Example University",
      "location": "Lyon",
      "start": "2020",
      "end": "2022",
      "description": "Distributed systems specialization.",
    },
  ],
  "spoken_languages": [
    { "language": "French", "level": "Native" },
    { "language": "English", "level": "Fluent (C1)" },
  ],
  "certifications": ["AWS Certified Cloud Practitioner (2024)"],
  "interests": ["Open source", "Climbing"],
  "preferences": {
    "target_roles": ["Backend Developer", "Software Engineer"],
    "locations": ["Paris", "Remote France"],
    "remote": "hybrid or full remote",
    "contract_types": ["Permanent (CDI)"],
    "seniority": "mid-level",
    "salary_min": "50k EUR gross per year",
    "availability": "Available immediately",
    "banned_keywords": ["PHP", "consulting firm"],
    "banned_title_keywords": ["internship", "apprenticeship"],
    "banned_companies": ["Evil Corp"],
    "notes": "No frequent business travel.",
  },
}

Anything not covered by these keys can be added freely: the AI reads the whole file.

The configuration (config.json)

The configuration is not sent to the AI. Unknown keys are rejected. The example below shows every key, with the default value of each one:

{
  "searches": [
    // one or more searches
    {
      "source": "linkedin", // see "Sources" below
      "keywords": "software engineer",
      "location": "Paris, Île-de-France, France",
      "posted_within": "day", // day | week | month | any
      "options": {}, // source specific settings, see "Sources"
    },
  ],
  "scraping": {
    "request_delay": 1.5, // seconds between two requests
    "cache_max_age_days": 7, // offer details older than this are downloaded again
    "max_results": 25, // offers listed per search, on every source
    "francetravail": {
      // settings of the "francetravail" source, see "Sources"
      "client_id": "", // identifier of your francetravail.io application
      "client_secret_env": "A_JOBSEEKER_FRANCETRAVAIL_SECRET", // environment variable holding its secret
    },
  },
  "matching": {
    "min_score": 60, // minimum score to select an offer, see "How it works" below
    "report_language": "English", // language of the reason given for each score
  },
  "templates": {
    "cv": "classic", // see "Templates" below
  },
  "ai": {
    "provider": "claude", // see "AI providers" below
    "path": null, // program path, defaults to the provider's program ("claude", "llm")
    "batch_size": 5, // offers evaluated per request
    "concurrency": 3, // requests running at the same time
    "timeout": 900, // seconds per request
    "max_attempts": 2, // attempts when the answer does not match the expected schema
    "claude": {
      // settings of the "claude" provider, used in this example
      "model": "sonnet", // see the "/model" list in the Claude Code interface
      "safe_mode": true, // ignores the user's CLAUDE.md, hooks, MCP servers...
      "max_budget_usd": null, // cost cap per request (float, e.g. 0.05, or null)
      "extra_args": [], // extra arguments passed to the "claude" program
    },
    "generic": {
      // settings of the "generic" provider, used when "--ai generic" is passed
      "args": [], // program arguments, "{prompt_file}" is replaced by the prompt file path
    },
  },
  "output": {
    "type": "stdout", // see "Outputs" below
    "email": {
      // settings of the "email" output, used when "--output email" is passed
      "smtp_host": "smtp.gmail.com", // SMTP server host
      "smtp_port": 587, // SMTP server port
      "username": "", // SMTP username, also used as the sender address
      "password_env": "A_JOBSEEKER_SMTP_PASSWORD", // environment variable holding the password
      "to": null, // recipient, defaults to identity.email from the profile
      "cv_base_url": "http://127.0.0.1:8000", // public URL of "a-jobseeker serve", used in the CV links
    },
  },
}

Sources

linkedin

Public (logged-out) LinkedIn job search pages, no account needed. location is free text, as typed on LinkedIn.

options are raw LinkedIn search parameters, added to the query string:

"options": {
  "geoId": 105015875, // LinkedIn location id, more precise than "location"
  "distance": 25, // kilometers around the location
  "f_WT": "2,3", // workplace: 1 on-site, 2 remote, 3 hybrid
  "f_E": "2,3", // experience: 1 internship, 2 entry, 3 associate, 4 mid-senior, 5 director, 6 executive
  "f_JT": "F,C", // job type: F full-time, P part-time, C contract, T temporary, I internship, V volunteer, O other
  "f_C": "1035" // company id
}

welcometothejungle

Public search index queried by the Welcome to the Jungle website, no account needed. The descriptions come from its public API.

The first comma separated part of location is matched against the office city, region or country code (Paris, Ile-de-France, FR).

options:

"options": {
  "language": "fr" // fr | en, language of the offer links
}

francetravail

Official France Travail "Offres d'emploi v2" API: the France Travail offers, and those of its partner job boards that agreed to share them (HelloWork, APEC, Cadremploi...). A free account is needed:

1. Create an account on francetravail.io and create an application.
2. Subscribe the application to the "Offres d'emploi v2" API (free).
3. Copy the application identifier to scraping.francetravail.client_id, and export its secret:

export A_JOBSEEKER_FRANCETRAVAIL_SECRET="..."

The first comma separated part of location is a French region, department or city name (Ile-de-France, Gironde, Nantes), resolved to its code with geo.api.gouv.fr. Paris, Lyon and Marseille are resolved to the codes the API accepts.

options are raw API parameters, overriding the ones built from the search:

"options": {
  "departement": "75,92", // up to 5 department codes, replaces "location"
  "commune": "44109", // INSEE city code, replaces "location"
  "region": "11", // region code, replaces "location"
  "distance": 20, // kilometers around "commune"
  "typeContrat": "CDI,CDD", // CDI, CDD, MIS (temping), SAI (seasonal), LIB (self-employed)...
  "experience": "2", // 1 less than 1 year, 2 from 1 to 3 years, 3 more than 3 years
  "tempsPlein": true, // true full-time, false part-time
  "qualification": 9, // 0 non-executive, 9 executive ("cadre")
  "salaireMin": 45000 // with "periodeSalaire": "A" (yearly), "M" (monthly), "H" (hourly)
}

The API returns at most 1150 offers per search: narrow the search (keywords, location, posted_within) rather than raising scraping.max_results beyond.

Indeed is not supported: its pages are protected against automated access, and it has no job search API.

AI providers

• claude (default): Claude Code in print mode, with native JSON schema validated output.

• generic: any command line AI. The full prompt, including the expected JSON schema, is written to the program's stdin, or to a temporary file when an argument contains {prompt_file}. The JSON object is extracted from its output.

# e.g. with the "llm" CLI (https://llm.datasette.io), which reads the prompt from stdin
a-jobseeker run --ai generic --ai-path llm

// e.g. a program taking the prompt as a file argument
// replace "{prompt_file}" with the path of the prompt file
"ai": {
    "provider": "generic",
    "path": "/usr/local/bin/my-ai",
    "generic": { "args": ["--input", "{prompt_file}"] }
}

Whatever the provider, answers are validated against the expected schema, and retried (max_attempts) with the validation errors when they do not match.

Outputs

stdout

Prints the text report shown in Quick start, listing every matching offer with its score, the reason of this score, its application link and the path of its CV. No setting.

email

Sends a single email listing every matching offer, with a CV link next to each Apply link that downloads that offer's cv.pdf. The run's logs are attached as run.log. Its settings are read from output.email (see The configuration).

The CVs are not attached: they are published by a-jobseeker serve, running on the same machine as the periodic run, which serves the PDF files of <data dir>/applications/ (no directory listing, application.json files are not served). cv_base_url is the address under which that server is reachable from where you read your emails, e.g. http://<VPS address>:8000 with a-jobseeker serve --host 0.0.0.0, or an HTTPS URL of a reverse proxy (nginx, Caddy...) forwarding to 127.0.0.1:8000. The CV links are not protected by a password: anyone who has one can download that CV.

The connection is always encrypted: implicit TLS on port 465, STARTTLS on any other port. The program only authenticates when username is set, and that address is then used as the sender.

With Gmail, the account password is refused: create an app password instead, and export it under the name given by password_env:

export A_JOBSEEKER_SMTP_PASSWORD="..."

a-jobseeker run --output email

Templates

A template describes what the AI must write, and how it is rendered as a PDF. The CV is always named cv.pdf.

The CV is written in the language of the offer, English or French for now. That language is the one declared by the job board when it provides it, otherwise it is detected from the offer text. An offer written in another language gets an English CV.

CV classic

Single column, designed to be parsed correctly by ATS (applicant tracking systems used by recruiters):

• no table, image or icon; selectable text in reading order;
• standard PDF fonts, conventional section headings ("Professional Experience", "Education"...);
• bullets drawn as shapes (no garbage glyph in the extracted text), links displayed in plain text;
• PDF metadata filled in (title, author, keywords);
• layout automatically tightened when it makes the CV fit on one page.

Sections: professional title and summary, experiences, projects, technical skills, education, languages, and optional additional sections (certifications, CTF...).

How it works

scrapers ──► filters ──► AI selection ──► PDF rendering ──► output

The sources are scraped at the same time, and the searches of each source one after the other. Before the detail of an offer is downloaded, it is skipped if it was already evaluated, if it is banned by the profile preferences, or if another search already found it. Duplicates are detected by offer id, and by company and title across searches and sources (case, accents and gender markers such as "(H/F)" are ignored). The scrape command applies the same deduplication.

For each batch of offers, a single request is sent to the AI. It contains the profile, the offers and the description of the selected CV template (writing guidelines + JSON schema of the expected content). Up to ai.concurrency requests run at the same time. Everything before the offers is identical in every request, so the AI provider can reuse its prompt cache.

For each offer, the AI returns, through a structured output validated against this schema: the decision, a score, a one sentence reason and, if the offer is selected, the content of the CV.

Each application is saved in <data dir>/applications/<date>/<source>-<company>-<title>-<id>/: the CV and application.json (offer + generated content, which can be edited and rendered again with a-jobseeker render).

Development

python -m venv .venv && make
.venv/bin/a-jobseeker init
.venv/bin/a-jobseeker run

# ...edit the code, then
make check # formats, lints, type checks, and runs the tests

make check fails when the tests do not cover every line and branch of the code.

To publish a new version on PyPI:

1. Update __version__ in src/a_jobseeker/__init__.py, and check the package with make dist.
2. Commit and push, then publish a GitHub release whose tag is the version prefixed by v (e.g. v0.2.0).
3. The Publish workflow runs the checks, builds the package and uploads it to PyPI.

• Every extension follows the same pattern: subclass the base class, then add it to the registry of the package. The user then selects it by name in the configuration.

• A CV template defines a pydantic content_model, which describes the content the AI must write, and whose JSON schema and field descriptions are sent to the AI, short instructions and a render method.

• To add a language to the CV, create a templates/locales/<code>.json file.

Disclaimer

Offers are collected from public pages and APIs of job boards whose terms of use may restrict automated access: use the tool reasonably (moderate volumes, delay between requests).