Import limits and large API workflows
File size, schema complexity, selected-tool and live-response limits for mcpmint browser processing.
Processing budgets
Specifications are limited to 5 MiB of UTF-8 content on every import path, including paste and Update spec. File imports accept JSON, YAML and YML. Portable project files have a separate 20 MiB limit because they also contain normalized models and tool settings.
An import can contain up to 10,000 operations. Each generated server can contain at most 500 selected tools, including compact mode. Default and bulk selection stop at 500; split larger exports into separate servers. Structural budgets also reject deeply nested or excessively expanded schemas before they exhaust browser memory.
Processing tasks run in a worker, show their current stage and can be cancelled. A task that exceeds 60 seconds is terminated. The editor displays 100 endpoints per page, while filtering and selection presets operate on the full matching catalog.
Compact mode and large catalogs
Compact mode exposes list, schema lookup and invocation meta-tools instead of sending every selected tool definition to the AI client at once. It reduces the tool context presented to a model; it does not bypass import or export workload limits.
Local CLI live tests time out after 10 seconds and stop reading after 256 KiB of response bytes. Browser inspection and mocks send no API requests. Use the generated server to verify edited tools and authentication.
When an import fails
Reduce a very large specification to the operations you need, bundle external references locally, and remove unnecessary examples or deeply expanded schemas. For an unreadable or damaged file, export it again from its original source. A visible import error leaves the previous working project available.
Memory and storage availability depend on the browser and device. Save a portable project backup before replacing a specification, clearing site data or moving to another production domain.