Essay grading API, MCP server and Claude skill

Three ways in

All three use the same jv_live_ key for the essay grading API, issued after a Google sign-in at API access, and spend the same balance. The grading underneath is Jev, TypeSafe AI's judgment model: it answers typed questions with a probability for every possible answer and writes no text, so every score comes with the distribution behind it.

The essay grading API

The essay grading API is one endpoint: POST /api/v1/tools/grade grades one essay, or a batch, against a rubric. Choose essay for class essays (six criteria and a letter grade), common-app for personal statements, sat for the SAT School Day essay, or custom for your own rubric.

curl https://jev-agent.com/api/v1/tools/grade \
  -H "Authorization: Bearer $JAGENT_API_KEY" \
  -H "content-type: application/json" \
  -d '{"rubric": "essay", "level": "high school",
       "assignment": "Should high schools start later?",
       "essay": "…"}'

The response for our calibration essay of middling quality, trimmed:

{
  "result": {
    "percent": 78, "letter": "C+",
    "verdict": { "label": "B or higher", "yes": 0.31 },
    "criteria": [
      { "name": "Thesis and focus", "level": "Clear, specific thesis the essay mostly follows", "index": 3.1, "levels": 5 },
      { "name": "Evidence and support", "level": "Some relevant support, thinly developed", "index": 1.58, "levels": 5 }
    ],
    "checks": [{ "label": "Length", "status": "warn", "detail": "209 words — short for most essay assignments…" }],
    "stats": { "words": 209, "paragraphs": 5, "readingGrade": 4.5 }
  },
  "usage": { "credits_charged": 2, "credits_remaining": … }
}

index is the essay's exact position on the criterion's ladder, 0 at the bottom; the fraction says which way it leans. verdict.yes is a probability, and checks are counted facts that never feed the grade. Limits: 100 to 20,000 characters per essay, and for batches up to 30 submissions, eight criteria and a 3- to 6-point scale.

Batch grading on your own rubric

{
  "rubric": "custom",
  "custom": { "points": 4, "criteria": [
    { "name": "Claim", "description": "States a clear, arguable claim" },
    { "name": "Evidence", "description": "Supports the claim with specific evidence" }
  ]},
  "submissions": [{ "name": "A", "text": "…" }, { "name": "B", "text": "…" }]
}

Each submission is graded in its own call, five at a time, so a paper's score depends only on the paper and the rubric. Three essays came back together in 0.54 seconds in our test.

The resume check API

POST /api/v1/tools/resume-check takes resume and an optional job_description, and returns a 0–100 match score, the posting's keywords found and missing, ATS format checks, the level fit and the probability a recruiter would move the resume to interview — the same report as the AI resume checker page. A resume plus a posting is about 1,100 input tokens, or 2 credits.

The MCP server

The server speaks Streamable HTTP at https://jev-agent.com/api/mcp. Listing its tools needs no key, so a client can see what is offered before anyone signs up; calling a tool needs the key in an Authorization header. In Claude Code:

claude mcp add --transport http jagent https://jev-agent.com/api/mcp \
  --header "Authorization: Bearer $JAGENT_API_KEY"

In Cursor's mcp.json (VS Code's file is the same idea under a servers key, with "type": "http"):

{ "mcpServers": { "jagent": {
    "url": "https://jev-agent.com/api/mcp",
    "headers": { "Authorization": "Bearer jv_live_…" } } } }

Clients that only run local servers can reach it through the mcp-remote bridge, passing the same header. The seven tools:

Every tool returns a one-line summary for the model to read, followed by the full result as JSON with the credits charged. compare_versions asks its question twice with the order swapped, because on close calls the model favors whichever option it read second — 0.90 in one order and 0.53 the other way round for the same pair, in our tests.

The Claude skill

The jagent-graders skill is a single SKILL.md: when to reach for the graders, the two REST calls, how to read a level and an index, and what not to ask for. In Claude Code, unzip it into ~/.claude/skills/ and set JAGENT_API_KEY in the environment; in the Claude apps, upload the zip where skills are managed. Claude then calls the API with curl whenever a conversation turns to grading an essay or checking a resume, and does the rewriting itself.

Billing and errors

The essay grading API bills like the rest of the site: one credit buys 1,000 input tokens, rounded up with a minimum of one per call, charged after the call succeeds. A 400-word essay graded on the six-criterion rubric is about 1,300 tokens — 2 credits; each extra submission in a batch adds its own tokens. There is no output-token charge, because the model produces none.

What it does not do

It does not rewrite, summarize or generate feedback prose; it grades, and your code or your agent writes. It does not detect plagiarism or AI-written text, and its grades are estimates from written rubrics, not official scores. It keeps nothing: request bodies are scored and discarded, and the call log records the endpoint and the credits, never the text.

Questions