jobsync-mcp

Agentic job-aggregation MCP server. Any MCP client (Claude Desktop, Claude Code, Cursor) loads it, and the client's own model drives discovery via web_search + web_fetch. This server exposes deterministic helpers — filters, classifiers, SQLite dedup, Airtable / Markdown sinks, profile I/O — plus prompt templates that encode the full scraping workflow.

Model-agnostic. You bring the LLM; we handle the plumbing.

Requirements

• Node >= 22.5 (uses built-in node:sqlite)
• An MCP client with web_search and web_fetch tools (Claude Desktop / Claude Code have these)
• Optional: an Airtable Personal Access Token — or skip Airtable entirely and use the markdown sink

Quick start

# 1. Configure (interactive wizard — uses npx, no global install needed)
npx -y jobsync-mcp@latest init

# 2. Parse your resume
npx -y jobsync-mcp@latest onboard --resume /path/to/resume.pdf

# 3. Register with your MCP client (see below)

# 4. In your client, invoke the onboarding prompt, then the scrape workflow

Step 2 — jobsync-mcp init

The wizard walks through:

Sink

Where should new jobs land?

• airtable — your own Airtable base (full schema, queryable)
• markdown — append to a local markdown file (no Airtable account needed)
• both

If you picked Airtable

Create a PAT at https://airtable.com/create/tokens with these scopes:

• data.records:read
• data.records:write
• schema.bases:read
• schema.bases:write (only if you want jobsync to create a new base for you)

The wizard then asks:

• Do you already have a base?
  • Yes → paste its base ID (app…)
  • No → paste your workspace ID (wsp…, visible in the Airtable URL when viewing a workspace), and the wizard creates a new JobSync base with the full Jobs table schema for you.

Other options

• Markdown output path — default ~/.jobsync/jobs.md
• Lookback hours — how fresh jobs must be (default 12)
• US-only filter (default yes)
• Enable ATS fast-path fetchers — direct job-board API calls for high-volume reliable pulls (default off; the agentic path is the default)
• Branded tool output — prefix every jobsync tool response with the Coral Labs marker (🪸 jobsync · Coral Labs). Default on; set brandedOutput: false in ~/.jobsync/config.json to silence it.

Config is written to ~/.jobsync/config.json.

Step 3 — jobsync-mcp onboard --resume PATH

Parses your resume (PDF / DOCX / TXT / MD) and saves the raw text to ~/.jobsync/profile/raw-resume.txt. The structuring into skills.md, experience.md, projects.md happens in Step 5 via the MCP prompt — driven by your client's model.

Step 4 — Register with your MCP client

Claude Code (recommended)

claude mcp add --scope user jobsync npx -- -y jobsync-mcp@latest server
claude mcp list        # expect: jobsync: npx  - ✓ Connected

The --scope user flag makes it available in every project. No global install needed — npx always pulls the latest published version when Claude Code starts.

Claude Desktop / Cursor

Add this block to your config file manually:

{
  "mcpServers": {
    "jobsync": {
      "command": "npx",
      "args": ["-y", "jobsync-mcp@latest", "server"]
    }
  }
}

• Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows)
• Cursor — ~/.cursor/mcp.json

Restart the client.

Step 5 — Run the prompts inside your client

Onboarding (one time)

In a Claude Code session:

/mcp__jobsync__onboard_profile

The agent will:

1. Read raw-resume.txt
2. Write skills.md, experience.md, projects.md via profile_write_file
3. Propose 3–7 target roles and ask you to confirm / add custom / exclude any

Edit any file in ~/.jobsync/profile/ by hand later — the agent re-reads them every run.

Primary workflow

/mcp__jobsync__scrape_jobs_workflow

Optional args:

• companies=Anthropic,Stripe — prioritize these
• roleKeywordsOverride=SDE,Backend Engineer — one-off override, ignoring profile
• lookbackHours=48 — widen/tighten the recency window

The agent runs: profile_read → web_search → web_fetch → classify_job_batch → cache_is_seen (+ airtable_list_recent_jobs fallback) → airtable_upsert_job or markdown_append_jobs (per your sink) → cache_mark_seen.

~/.jobsync/ layout

~/.jobsync/
├── config.json
├── cache.db             # SQLite dedup cache
├── jobs.md              # markdown sink (if enabled)
└── profile/
    ├── skills.md        # editable
    ├── experience.md    # editable
    ├── projects.md      # editable
    ├── roles.json       # { detected, custom, excluded }
    └── raw-resume.txt

Tools

| Tool | Purpose | |---|---| | jobsync_ping | Sanity check | | filter_us_location | Drop non-US postings | | filter_title_keywords | Keyword match on titles | | detect_industry_tags | Industry + FAANG+ / YC / H1B tags | | classify_job_batch | All filters + tags in one call | | airtable_upsert_job | Batch upsert into user's base | | airtable_list_recent_jobs | Dedup fallback | | airtable_get_schema | Inspect base/table + active config | | airtable_list_bases | List bases the PAT can access | | airtable_create_base | Create a new base with the JobSync schema | | markdown_append_jobs | Append jobs to the markdown log | | cache_is_seen | SQLite dedup lookup | | cache_mark_seen | Mark as seen after upsert | | cache_prune | TTL-based cache trim | | profile_read | Load skills / experience / projects / roles | | profile_write_file | Overwrite a profile markdown file | | profile_update_roles | Mutate detected / custom / excluded | | profile_parse_resume | PDF / DOCX / TXT → raw text | | fetch_greenhouse_jobs | ATS fast-path (flag-gated) | | fetch_lever_jobs | ATS fast-path (flag-gated) | | fetch_ashby_jobs | ATS fast-path (flag-gated) | | ashby_get_date_posted | Scrape real datePosted from an Ashby job page's JSON-LD | | ashby_get_date_posted_batch | Batch version with optional lookbackHours filter | | verify_job_link | Check if a job URL still resolves to a live posting | | verify_job_link_batch | Batch link verification (max 20) |

Prompts

• onboard_profile — one-time resume structuring + role selection
• scrape_jobs_workflow — primary loop
• extract_job_fields — per-posting extraction helper

Resources

• prompts://extraction-schema — field schema the extraction prompt targets

CLI

jobsync-mcp                         Start the MCP server (stdio)
jobsync-mcp init                    Interactive config wizard
jobsync-mcp onboard --resume PATH   Parse a resume (.pdf / .docx / .txt / .md)
jobsync-mcp print-client-config     Emit MCP client JSON snippet
jobsync-mcp status                  Show config path + existence

Running as a cron job

The server is just a stdio MCP peer — scheduling is the client's job. With Claude Code, drop this in your crontab / Task Scheduler:

claude -p "invoke the scrape_jobs_workflow prompt from jobsync with lookbackHours=24"

Troubleshooting

• node:sqlite not found — you're on Node <22.5. Upgrade Node.
• claude mcp list shows "Not connected" — verify jobsync-mcp is on PATH (where jobsync-mcp / which jobsync-mcp). Re-register with the absolute path if needed.
• Airtable 403 on airtable_create_base — PAT is missing the schema.bases:write scope. Regenerate.
• 0 new jobs on every run — search indexes lag; your lookbackHours window is too tight. Widen to 48–72, or enable enableFastPath and use fetch_ashby_jobs / fetch_greenhouse_jobs / fetch_lever_jobs against known company slugs.
• "Unknown skill" for slash command — your /mcp__jobsync__... form isn't supported in your Claude Code version. Either select the prompt from the /mcp menu, or just ask in plain English: "run the onboard_profile prompt from jobsync."

License

MIT