Connecting Claude & agents

Jobably speaks MCP over Streamable HTTP at https://mcp.jobably.com/mcp, authenticated with a personal access token.

1. Create a token

App → Settings → Personal access tokens. Give your agent the scopes it needs — read for browsing, add write if it should move cards, rate, or queue tailoring. The token is shown once.

The easy way: the moment a token is created, the Connect your AI assistant card right below shows the complete one-line command — Claude Code, Codex, or GitHub Copilot; Windows, macOS, or Linux — with your new token already substituted in. Copy, paste, done. The snippets below are the same commands for reference.

2. Connect your client

Claude Code (CLI)

claude mcp add jobably --transport http https://mcp.jobably.com/mcp \
  --header "Authorization: Bearer YOUR_TOKEN"

Claude Desktop / claude.ai (custom connector)

Add a custom connector with:

• URL: https://mcp.jobably.com/mcp
• Authorization header: Bearer YOUR_TOKEN

Generic MCP client (JSON config)

{
  "mcpServers": {
    "jobably": {
      "type": "http",
      "url": "https://mcp.jobably.com/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

3. Economy for agents

Every tool description declares a cost class — [free] or [metered: …] — so agents can be economical without guessing. Anything that spends quota or changes state additionally requires confirm: true in the arguments; a call without it returns an instructive error instead of acting.

See the tool reference for the full list.

4. Passing arguments correctly

Arguments are a plain JSON object. Two calls, shaped exactly right, cover the mistakes that actually happen:

{
  "name": "add_accomplishments",
  "arguments": {
    "bank": [
      "Negotiated $4.8M savings, 20% on $24M of capital equipment spend.",
      "Cut sourcing cycle time 38% across a $40M program."
    ]
  }
}

bank is a real JSON array — not a quoted string containing one — with one accomplishment per element. Commas stay INSIDE an entry: the first element above is one accomplishment, not three. Never pre-split text on commas; split on line breaks.

{
  "name": "create_search",
  "arguments": {
    "name": "VP Engineering — remote, US",
    "query": "\"VP Engineering\" OR \"VP of Engineering\"",
    "location": "United States",
    "experienceLevel": "5-10,10+",
    "excludeTitles": ["recruiter"],
    "remoteOnly": true,
    "cadenceHours": 24,
    "confirm": true
  }
}

confirm is the boolean true, not the string "true". Numbers may arrive as strings and coerce ("24" means 24), but unparseable values error loudly rather than being replaced with a default. Ids (jobId, searchId, operationId, artifactId, resumeId) always come from a prior call — never invent one.

Every tool in the reference carries an example call validated against the live schema in CI — copy its shape.

What an agent can actually do

The surface covers the whole product, not just search:

• Find work — one-off searches, and count_jobs sizes a query for zero job credits before anything is spent.
• Run searches on a schedule — create, edit (every filter, merged so nothing you did not mention is lost), pause, resume, run now.
• Diagnose a quiet search — get_search_health returns the compiled query the supplier actually received, which is what separates “nothing matched” from “this query cannot match”.
• Work the pipeline — matches with written rationales, board moves, notes, ratings, applications made elsewhere.
• Generate and review resumes — queue a tailor, poll it, read the produced document back as text, and read the gap list of what the posting wanted that the record does not evidence. An agent that cannot read the draft cannot help you check it.
• Improve future drafts — add to the accomplishment bank tailoring draws on, and mark style exemplars.

Deliberately absent: sign-in, personal access tokens, API keys, billing and account deletion are session-only, so an agent can never mint credentials or spend money outside the metered tools.

Searching for someone else

search_jobs runs one-off searches that are never scored against your profile and never saved. With forSomeoneElse: true, level and size filters are relaxed entirely — useful when a friend asks “can you look for staff-level roles for me?” without your rubric contaminating their results.