Model connection

Eregion reads .env in the project folder. Variables set in your shell override it. Restart the server after changing settings.

The model must support tool calling and structured output. Visual browser backup also needs image input.

Job sources

Searches use public company boards and, once you set it up, LinkedIn. No paid search key is needed.

Set up LinkedIn

1. Install the pinned mcp-server-linkedin==4.27.0 in its own environment
2. Copy docs/examples/linkedin-mcp.json, adjust the paths, and point MCP_CONFIG_PATH at it in .env
3. Sign in once with the server's --login mode in a separate browser profile, never your everyday one

Jev fast fit scoring

Optional. With Jev on, fit scoring uses Typesafe's Jev decision model alone, batches up to sixteen jobs, and caches decisions on your computer. Conversation and writing still use your main model.

TYPESAFE_API_KEY=your-typesafe-key
TYPESAFE_BASE_URL=https://api.typesafe.ai
TYPESAFE_DEFAULT_MODEL=jev-latest
TYPESAFE_ENABLED=true
TYPESAFE_CAREER_ENABLED=false

• Decisive answers are cached for 24 hours, uncertain ones for 5 minutes
• Undecided requirements stay unresolved and the score becomes a range
• No fallback to the main model: if Typesafe fails, scoring fails
• Sends extracted resume text, the job description and scoped facts. Inputs over 60,000 bytes are rejected, not truncated

Browser and workers

These are read from the shell environment, not .env. Set them before running migrations and the server.

# macOS / Linux: headless Chrome, two workers
export JOB_AGENT_HIDDEN_BROWSER=headless
export JOB_AGENT_APPLY_WORKERS=2
python -m uvicorn eregion.api.main:app --host 127.0.0.1 --port 8000

Data and privacy

The local database holds your details, conversations, evidence, application state and model-usage records. Treat it as personal data, and stop the server before copying it as a backup.

• The server listens on 127.0.0.1 only. It is a single-user app with no sign-in
• Model requests contain the resume, job, answer or document content a task needs
• On Windows, site logins are kept with DPAPI encryption, scoped to you and the site
• Screenshots reach your model only when visual browser backup is on

Good to know

Some sites need you

Login walls, unusual widgets, MFA and CAPTCHA can pause the agent.

Scores describe evidence

A match score reflects your resume, not an employer's decision.

Review before you send

Materials are checked against sources; read them anyway.

Text PDFs only

Scanned resumes need text extraction first.

Evaluation

Eregion is tested with four synthetic applicants: a junior data analyst in New York, a senior backend engineer for remote US, an ML intern in California, and a junior UX designer in Austin. Live runs have fixed budgets and never apply or create schedules. A person reviewed one full set of results by hand.

Shown jobs

0 of 23 were wrong in the human review. The latest runs showed 123 jobs with 0 structural errors.

What to show

28 of 29 labelled postings were shown or hidden the way the reviewer wanted.

Tiers

About two in three match the reviewer; most misses are one tier apart.

Known limits

The agent often stops searching early and offers to continue. Scoring accuracy against human labels is not measured yet.

The evaluation report summarises everything that was tested, including what failed along the way.

Updating

Stop the server and back up your data, then from the project folder:

git pull --ff-only
python -m pip install -r requirements.txt
python -m alembic upgrade head
python -m uvicorn eregion.api.main:app --host 127.0.0.1 --port 8000

Reload the Eregion Lens extension after updating. Never delete the database to fix a migration error.

How it works inside

One shared agent loop runs every autonomous task: chat, scheduled searches and applications. It chooses tools and the order of work. Code checks what the model cannot be trusted to judge alone: quoted evidence must appear in the posting, budgets are enforced, and approvals and the final submit stay with you.

Read the architecture notes for the loop, context and persistence details.

Development

python -m pytest -q
node --test "tests/*.mjs"