Indraft

What the Indraft MCP server does

The Indraft MCP server gives an AI agent access to a hosted CRM designed for agent-driven work. The default data model includes companies, contacts, opportunities, tasks, and interactions, so an agent can begin working without first designing a schema. It can also work with custom fields and user-defined object types when the business needs a different model.

The system records how each value was established, such as through an import, an observation, an inference, or a human-provided or human-confirmed statement. These assertion types have different standing. A weaker assertion cannot silently replace a stronger one, and a write that cannot be applied is reported as held rather than presented as successful.

Changes include the actor, the entry surface, and a field-level difference. Most changes can be reversed. The service also avoids ambiguous operations: cross-currency totals require a stated conversion rate, and bulk deletion is not available.

How it works

Clients connect to the hosted MCP endpoint at https://api.indraft.io/mcp. The connection uses Streamable HTTP and OAuth 2.1. After authorization, the agent can inspect the CRM schema before making changes. This lets it learn available object types, limits, and permitted operations instead of assuming field names or capabilities.

The Indraft MCP server exposes 45 tools, subject to the permissions of the credential in use. Tool metadata includes a readable title plus read-only, destructive, and open-world hints, allowing an MCP client to distinguish reads from writes before execution. Aggregations over multiple currencies return separate currency groups when no conversion rate is available rather than producing a misleading combined total.

Setup and configuration

Use the hosted endpoint rather than running a local process. In Claude, add the endpoint with the HTTP transport; authorization opens in a browser. Claude Desktop and claude.ai can add the same URL as a custom connector. ChatGPT supports it through Settings, Connectors, and Add custom connector, followed by OAuth authorization.

Cursor, Windsurf, VS Code, and other clients can connect if they support Streamable HTTP and OAuth 2.1. After connection, an appropriate first instruction is to ask the agent to read the CRM schema, explain how it would model the business, and wait for approval before writing records.

No server environment variable or API key is specified for the hosted MCP connection. The repository also includes a TypeScript SDK for REST API use; its example accepts an INDRAFT_TOKEN, but that is for SDK clients rather than the hosted MCP setup.

Tools and capabilities

The available tool set covers:

• Reading and writing CRM records
• Creating or modifying schema elements
• Merging records
• Importing data in bulk
• Aggregating CRM data
• Tracking assertion provenance and field-level changes
• Reporting held writes when a value cannot override a stronger assertion
• Returning recovery hints for typed SDK errors

The broader repository also documents a REST API with an OpenAPI description, a TypeScript SDK, and export options for workspace data. Those interfaces are separate from the MCP connection but may suit services that need direct REST access.

Limitations and notes

Permissions determine which tools a credential can call. Standing overrides may be refused by workspace policy, in which case the write must be retried without the override or reviewed by a person. Cross-currency totals need an explicitly stated rate; without one, results remain grouped by currency. The hosted service has separate terms and privacy policies from the repository's MIT license.