API & MCP Access

The foundation is API-first, so everything you do in the Collections UI is available programmatically, through three interfaces:

A REST API for applications and scripts, fully documented with an interactive playground
An MCP endpoint (JSON-RPC 2.0) for AI agents and MCP-compatible clients
The AI Collection App, a reusable Builder app that exposes the API as ready-made instructions inside your own workspaces (see In the Builder below)

Collections API Reference

Full endpoint documentation with request/response schemas and a live playground: collections CRUD, records, analytics, and the MCP endpoint.

REST API at a Glance

Area	What you can do
Collections	List (paginated, optional estimated row counts), create (with a typed schema), read, update metadata and sharing, delete
Records	Query with MongoDB-style filters, insert in batches (up to 1,000), bulk-insert for large imports (up to 5,000 per call), read, update, upsert, delete
Analytics	Count, distinct values, aggregations (sum, avg, count, group by)
Schema	Add columns and apply indexes or unique constraints after creation
AI	Suggest a schema from a description, ask a plain-language question about a collection
Sharing	Grant, list, change, and revoke access for users, groups, or the whole org
API Keys	Mint, list, rotate, and revoke org API keys scoped to a collection

Queries accept MongoDB-style filters:

{
  "status": "active",
  "deal_value": { "$gte": 50000 },
  "stage": { "$in": ["Proposal", "Negotiation"] }
}

For safety, operators that execute code or cross-reference data ($where, $function, $expr, $lookup, $merge, $accumulator, $out) are rejected.

MCP Endpoint

Agents and MCP clients connect to the workspace webhook endpoint (shown in the Settings page of the product):

POST /v2/workspaces/slug:ai-collection-v3/webhooks/ai-collection/mcp

Protocol: JSON-RPC 2.0 (initialize, tools/list, tools/call). The endpoint exposes the data tools agents use:

Tool	Purpose
`data_discover`	Return the schema, row count, and a few sample rows (call this first)
`data_create_collection`	Create a named collection, optionally with a typed schema
`suggest_schema`	Propose a typed schema from a plain-language description
`data_insert`	Insert one or more records
`data_query`	Search records with filter, sort, and pagination
`data_get`	Fetch a single record by ID
`data_update`	Update records matching a filter
`data_upsert`	Insert-or-update on matching fields
`data_delete`	Delete records (bulk delete requires explicit configuration)
`data_count`	Count matching records
`data_distinct`	Unique values of a field
`data_aggregate`	Sum / average / count with grouping

Tool arguments mirror the REST payloads; see the API Reference for the schemas. To bind an agent to a single collection, point it at the per-collection endpoint instead; the collection_id is then implicit on every call:

POST /v2/workspaces/slug:ai-collection-v3/webhooks/v1/collections/{collection_id}/mcp

Text filters match exactly and are case-sensitive: informatique will not match a stored Informatique. For case-insensitive matching use a regex such as {"category":{"$regex":"^informatique$","$options":"i"}}, or call data_distinct first to get the exact value. The ask tool and endpoint handle this for you.

Schema Format

When creating a collection programmatically, columns are declared as:

{
  "name": "products",
  "properties": {
    "product_name": { "type": "text" },
    "price": { "type": "double" },
    "stock": { "type": "integer", "nullable": true },
    "active": { "type": "boolean" },
    "released_at": { "type": "date" },
    "specs": { "type": "json" }
  }
}

Supported types: text, integer, double, boolean, date, datetime, json, array.

In the Builder (App)

For workflows built in Prisme.ai Builder, you do not need to call the REST API with raw fetch. Import the AI Collection App and call its instructions directly, the same way you would use any other app.

Install the app

In your workspace, open Apps, add AI Collection App, and (optionally) set an API key in its configuration. When an automation runs in a user context, the user’s identity is propagated automatically, so a key is only needed for headless flows (crons, event listeners).

Call an instruction

Each REST endpoint is exposed as a clean instruction, grouped into folders in the editor:

Folder	Instructions
Collections	`listCollections`, `createCollection`, `getCollection`, `updateCollection`, `deleteCollection`
Data	`insertData`, `bulkInsert`, `queryData`, `countData`, `distinctData`, `aggregateData`, `upsertData`, `getRecord`, `updateRecord`, `deleteRecord`
Schema & AI	`updateSchema`, `addIndexes`, `suggestSchema`, `ask`
Charts	`getCharts`, `saveCharts`
Sharing	`listAccess`, `share`, `revokeAccess`
API Keys	`listApiKeys`, `createApiKey`, `deleteApiKey`, `rotateApiKey`

For example, to create a collection and query it from an automation:

- ai-collection.createCollection:
    name: leads
    properties:
      email: { type: text }
      score: { type: integer }
    output: created
- ai-collection.queryData:
    collection_id: '{{created.collection_id}}'
    filter:
      score: { $gte: 80 }
    output: hotLeads

Every instruction returns the same payload as its REST counterpart, and forwards errors as { error: true, status, message } so you can branch on failures.

Limits

Limit	Value
Batch insert	1,000 records per call
Bulk insert	5,000 records per call (all-or-nothing per batch)
Query page size	100 records max per call (default 20); paginate for more
Aggregation steps	10 per call
Distinct values	1,000 max returned
Saved charts	50 per collection
Collection name	64 characters, alphanumeric plus `-` and `_`
Rate limits	100 requests/min for reads and most operations; 20/min for inserts; 120/min for bulk inserts; 50/min for upserts

Behavior Notes

Protected fields: record IDs, creation timestamps, and ownership metadata are managed by the platform; they are stripped from update payloads.
Bulk deletes are opt-in: deleting multiple records in one call is rejected unless enabled in the platform configuration.
Duplicate names are rejected: collection names are unique per owner.
Audit events: every write (insert, update, upsert, delete, collection created) emits an audit event with the actor and target collection.

Overview

Chat

Agent Creator

Knowledges

Collections

Builder

Governe

Engage

Insights

Collections API Reference

REST API at a Glance

MCP Endpoint

Schema Format

In the Builder (App)

Limits

Behavior Notes

Next Steps

Collections API Reference

Using Collections with Agents

Collections API Reference

​REST API at a Glance

​MCP Endpoint

​Schema Format

​In the Builder (App)

​Limits

​Behavior Notes

​Next Steps

Collections API Reference

Using Collections with Agents

REST API at a Glance

MCP Endpoint

Schema Format

In the Builder (App)

Limits

Behavior Notes

Next Steps