Quickstart

1. Connect your agent

Add the MCP endpoint to Claude, ChatGPT, or any MCP client:

https://api.indraft.io/mcp

Your client will send you through an authorization flow. Nothing is pasted, and you can revoke the connection later from the workspace.

2. Give it this

Set up Indraft for my business. Read its CRM schema first, then tell me what you think we do and how you plan to model it, and wait for my answer before you write anything. Configure the pipeline to match how we actually sell, add the custom fields that earn their place, and if something we run on is not a company, contact, opportunity, interaction or task, define an object type for it rather than forcing it into a field. Then populate what you already know from our conversations. Do not create duplicate people or companies: if an upsert comes back ambiguous, tell me rather than picking one.

This prompt ships with the product and is versioned in the repository rather than written into this page, so what you copy here is what the product tests against.

3. Watch it work

The activity view fills as the agent writes. Every change carries the actor that made it, whether it was observed or inferred, and which surface it came through, so you can tell an agent's claim from a colleague's correction at a glance.

What a first session costs

Nothing, and nothing stops halfway. Every new workspace runs on the top plan's limits for seven days, so an agent can configure the pipeline, add the fields it needs, and populate months of history without meeting a number. No card.

After the week you fall back to the plan you are on. Free holds 2,500 records and 10,000 changes a month, which is comfortably more than a founder-operator with a couple of dozen customers needs for the backfill and then some. If you end up over, writes are refused and nothing else changes: everything stays readable, correctable, and exportable, and archiving what you are not working brings you back under without deleting anything.

There is no capability behind a paywall: merge, undo, provenance, bulk writes, webhooks, export, and MCP are on every plan, because those are what the product is. Paid plans raise the numbers.

One limit that is per call rather than per month: a bulk request carries at most 100 operations. A population run makes several of them.

If something is refused

Read the recoveryHint. Every failure carries one, written for an agent rather than for a human reading a log, and it says what to do rather than what went wrong. An ambiguous upsert, for example, returns candidates and writes nothing: Indraft does not guess which of two people you meant.