Skip to main content
Nimt MCP is continuously evolving. We add capabilities and improve how tools communicate with AI assistants. The current service is in pre-release: tool names, parameters, response structures, and behavior may change, and there is no fixed-schema or backward-compatibility guarantee during this period.

​
If you use ChatGPT or Claude

Keep your client updated. If it shows an outdated tool list, refresh the Nimt connection or reconnect using the client guide. Review newly requested permissions before approving them. You do not need identical wording to use Nimt successfully. Your assistant chooses how to summarize the data, so formatting, phrasing, and the order of an AI-written answer can differ between conversations. Read the changelog for documented changes to the customer MCP.

​
If you build a custom integration

  • Discover the current tool catalog and input/output schemas with tools/list. Do not permanently cache a copied tool definition.
  • Validate inputs against the current schema. Do not send unknown parameters because a client or another Nimt interface accepts them.
  • Consume structured results according to the advertised output schema. Avoid parsing the assistant’s prose or comparing entire serialized responses to an example copied from these docs.
  • Handle additional output fields without assuming your stored schema is the only possible shape. Validate the fields you rely on; do not silently ignore missing required fields or changed types.
  • Treat IDs as opaque values, paginate explicitly, and compare records by ID rather than array position unless an order is documented.
  • Handle missing tools, authorization failures, and changed schemas explicitly. Refresh discovery and review the changelog instead of blindly retrying.
The server version reported during MCP initialization identifies the deployed implementation. It is not a version-pinning option for this hosted endpoint. If your integration requires a fixed contract, contact support@nimt.ai before depending on it for an unattended or business-critical workflow.

​
Live data and exact equality

Even without a tool change, two calls can return different results:
  • Projects, configuration, membership, and stored data can change between calls.
  • fetchedAt changes with each metadata collection.
  • Paginated calls and the fields in a metadata response are not a single transactional snapshot of the entire workspace.
  • Lists without a documented ordering should not be compared by position.
For comparisons, align the workspace, project, filters, pagination, and collection period. Compare the relevant values by identifier and account for volatile fields such as fetchedAt. This is not a reason to dismiss incorrect results. A tool result should match its advertised schema and accurately describe the data it reads. If comparable results conflict or a required field is missing, report the discrepancy through troubleshooting.

​
How we communicate changes

We favor additive changes and preserving existing behavior where practical. Changes that require customer action should have a changelog entry explaining what changed and how to adapt. Security fixes may require prompt changes or reauthorization. During pre-release, do not assume a fixed notice period, a pinned tool set, or that every schema change will happen automatically inside an existing chat. Refresh the connection after a relevant update and retest workflows you rely on.