Get started

AI skills & agents

Give Claude, Cursor, ChatGPT and custom agents MessyAPI tools.

MessyAPI exposes eight high-level AI skills. An agent sends a plain-language goal plus records (or text); MessyAPI builds a plan from operations that actually exist, runs it, and returns the plan, per-step report and result. Anything it can't do comes back as unsupported_capability — it never guesses.

Quickstart (REST)

curl https://messyapi.com/api/v1/ai/clean \
  -H "Authorization: Bearer $MESSYAPI_KEY" -H "Content-Type: application/json" \
  -d '{"goal":"normalize phones, validate emails, find duplicates, redact PII","records":[{"name":"jane DOE","email":"jane@example.com","phone":"(555) 123-4567"}]}'

Skills

POST /api/v1/ai/clean — messy_clean: normalize phones, addresses, URLs, names. Accepts multi-step goals.
POST /api/v1/ai/validate — messy_validate: emails, data quality; formats/schema with options.
POST /api/v1/ai/match — messy_match: duplicate and similar-record clusters (reports groups, doesn't delete rows).
POST /api/v1/ai/transform — messy_transform: CSV/TSV/JSON conversion, flatten; pivot/unpivot/rank via operations.
POST /api/v1/ai/protect — messy_protect: PII and secret detection, redacted copy.
POST /api/v1/ai/analyze — messy_analyze: profile, missing values, quality, statistics; text language/sentiment/keywords/entities/summary.
POST /api/v1/ai/batch — messy_batch: same plan over up to 10,000 records in 1,000-record chunks (duplicate detection runs per chunk).
POST /api/v1/ai/execute — Messy Execute (limited preview): send a goal or template (crm-import-prep, lead-validation, ai-ready-dataset, pii-safe-export, dedupe-contacts, data-profile) with up to 10,000 records; returns the plan, results and a verification report (PASS / PASS_WITH_WARNINGS / FAIL, record counts per step, lost fields, duplicates from the cluster step, PII scan done). Quality score = filled cells ÷ (records × fields) × 100.
POST /api/v1/ai/verify — Messy Verify (limited preview, free): check records against required fields and types (string, number, boolean, email).
GET|POST /api/v1/ai/discover — messy_discover: free, no key; returns relevant real operations for a goal.

Request body

{
  "goal": "string (optional)",
  "records": [ { ... } ],        // or
  "text": "string",
  "operations": [ { "operation": "pivot", "options": { ... } } ],  // advanced, optional
  "plan_only": true               // return the plan, free
}

MCP (Claude Desktop, Claude Code, Cursor, VS Code)

Point any MCP client at https://messyapi.com/mcp with Authorization: Bearer YOUR_KEY. Add ?mode=skills to list only the 8 skills plus 4 outcome tools; without it every operation is also listed.

claude mcp add --transport http messyapi "https://messyapi.com/mcp?mode=skills" \
  --header "Authorization: Bearer $MESSYAPI_KEY"

ChatGPT custom GPTs, Gemini, custom agents

These connect through the OpenAPI spec (/docs/openapi.json) as actions/function definitions, or call the REST endpoints above directly. This is a standard OpenAPI/REST connection, not a native plugin.

Permissions

Keys limited to categories (Clean, Validate, Match, Transform, Protect, Analyze) are checked against every step of the plan on the server. If any step needs a category the key lacks, the whole call is refused with 403 forbidden naming the step. The goal text can't override this.

Limits & billing

Skills meter like any endpoint: records × each step's weight, through the same quota, rate limits and spend alerts. plan_only and discover are free. Up to 1,000 records per call (10,000 for batch).

Security

Goal text is treated as data; prompt-injection markers are reported as injection_detected and never change what runs.
Planning is deterministic — no model picks operations, so plans are reproducible.
Skills never fetch URLs. Secrets found in data are always removed by redaction.
Request logs keep only metadata (skill, goal excerpt, step counts), never your records.

Try it

The live demo on /ai runs the same planner without a key (25 records max).