What is a CSV import API?

A CSV import API lets a backend or agent run the import pipeline without a UI: upload a file, get column-mapping proposals, run transforms as a dry run, validate, and apply the clean rows into the host system. Mildport exposes this as a versioned, license-keyed REST API described by OpenAPI.

Can AI agents use the import API?

Yes — Mildport ships an MCP server inside your deployment with tools like analyze_file, propose_mapping, dry_run_transform and execute_import. Agent access is licensed and entitlement-gated, agents answer ambiguity questions only with cited evidence, and high-stakes decisions can require human confirmation.

How do I keep an automated import safe?

Mildport keeps automation governed: deterministic matching first, typed ambiguity questions instead of silent guesses, per-import decision logs, and delivery over signed webhooks with retries. Executing an import as an agent requires both a license entitlement and an operator-enabled switch.

CSV import API — headless imports for backends and AI agents

Это содержимое пока не доступно на вашем языке.

A CSV import API runs the whole pipeline headlessly: ingest → map → transform (dry-run first) → validate → apply. Mildport’s API is versioned (/api/import/v1), license-keyed, described by OpenAPI, and runs in your infrastructure — so the automation and the data share a boundary.

The headless flow

# 1. Ingest a file (returns a record id)
curl -X POST $BASE/api/import/v1/ingest/file \
  -H "Authorization: Bearer $LICENSE" --data-binary @contacts.csv

# 2. Ask for mapping proposals (deterministic scores + reasons)
curl -X POST $BASE/api/import/v1/mapping/column \
  -H "Authorization: Bearer $LICENSE" -d @columns.json

# 3. Dry-run transforms, then apply — clean rows arrive on your signed webhook

Every response is explainable: match percentages, reason codes, calibrated probabilities. Ambiguity doesn’t fail the run or guess silently — it becomes a typed question your system (or a person) answers, and answered questions teach the per-tenant mapping memory.

Agent-ready, inside your boundary

The same deployment exposes an MCP server (/.well-known/mcp/server-card.json): analyze_file, propose_mapping, dry_run_transform, list_catalogs, and — behind a license entitlement and an operator-enabled switch — execute_import. Agents are first-class, auditable actors: their answers require cited evidence, and every decision records who decided.

Unlike cloud importers’ agent surfaces, this one fronts your own instance — an agent can import into your system without customer data ever leaving it.

Proof, not promises

Imports produce a decision log (mapping reasons, transforms, actors) and can produce a signed receipt that a third party verifies offline. Delivery is idempotent, signed and retried. See self-hosting for the deployment and pricing for the published rate card.

Prefer a UI? The same engine powers the embeddable widget.