Per-Provider Tuning Baselines

When you configure how your MTA shapes traffic to each destination, you need concrete values to start from. The values below come from the shaping files that the KumoMTA project maintains publicly. There are two files:

• the policy-extras/shaping.toml file maintained by the vendor, with defaults and the major providers (several sections are marked "provided directly by" the provider);
• the community/shaping.toml file contributed by the community, with more providers and automation rules driven by bounces.

Fetched 2026-07-20 from the main branch. These values drift as providers change their policies. Treat them as attributed starting points to check against your own deferral data, not as facts about current provider limits. See MTA Delivery Tuning for what each parameter means and how the automation model works.

How to read the tables:

• MX rollup: limits apply to each receiving MX infrastructure, so all recipient domains that resolve to the same set of MX hosts share one bucket. Provider blocks match by MX suffix. For example, anything with an MX under .google.com or .googlemail.com counts as "google", which covers Gmail and domains hosted on Google Workspace. According to the file's note of 2024-09-03, their MX hostnames resolve to common IP addresses, so they are shaped together.
• provider_connection_limit vs connection_limit: the provider variant is enforced across the whole provider rollup, and the plain variant applies to each site. Mimecast deliberately uses limits for each site, because each Mimecast site name is a separate regional fleet of MTAs.
• Where a cell is blank, the global default applies.

Global defaults

A default automation rule, for all destinations, matches the family of deferral messages that many providers use: Messages from <IP> temporarily deferred, All messages from <IP> will be permanently deferred, has been temporarily rate limited due to IP reputation, Unfortunately, messages from <IP> weren't sent, Server busy. Please try again later from. On a match, it sets max_message_rate = 1/minute and connection_limit = 1 for 90 minutes.

Provider baselines

Automated-throttle rules per provider (community-contributed)

The community file adds these adjustments on top of the baselines, in response to bounces. A trigger of N/hr means the rule fires once the response has matched N times in an hour, and every action expires after its duration. For what each response means, see the error codes for Gmail, Yahoo, Microsoft, Apple, Comcast and GMX and WEB.DE.

Reading the pattern

Across providers, the community's throttling logic follows the design principles in MTA Delivery Tuning:

• Responses about mechanical limits (too many connections, or too many messages per session) get a parameter cut for about an hour.
• Brief capacity problems ("server busy", "service unavailable") get short suspensions of 15–30 minutes.
• Reputation events (deferrals for user complaints, blocklist rejections) get the longest suspensions (2–4 hours), because continuing to send through them makes reputation worse.

Note also how low the baselines are compared with the global default. The biggest providers are shaped to 5 connections across the provider, and several regional providers (Orange, Yahoo Japan) to just 2.

Caveats

• The values are what worked for KumoMTA operators as of the fetch date. Providers change limits without notice, and the reputation of each IP address affects what a given sender actually gets. Check the values against your own deferral logs.
• Several entries carry dates in the file (Barracuda and Netvigator are marked 2026-03-27, and the note on the Google and Workspace rollup is dated 2024-09-03). Each row has its own freshness.
• The community file states explicitly that its rules are contributed as-is, and must be combined with the vendor's defaults file for full coverage.
• These are limits that senders impose on themselves, not guarantees published by providers, except where a row is marked "provided directly by" the provider (Yahoo, Comcast, Mimecast, mail.com, Mailgun).