By plan
Keep these capacity scopes distinct:
- The connection cap is per account, shared across every key. Five Pro connections are five in total, however many keys you spread them over. An indexer on one key, a staging box on another and a tab left open on a third are three of them. REST request quotas are also shared across account keys, but are separate from WebSocket capacity.
- The subscription cap is per connection. Each socket may hold that many distinct subscriptions; opening another connection (within the cap) gets you another set.
- A multi-wallet list is one subscription. A
…ByUserssubscription carrying 1,000 wallets spends one slot, not 1,000. The 1,000-wallet ceiling is the same on every plan; what differs between plans is how many such lists you may hold at once.
l4Booksubscriptions count against the subscription cap above like any other — there is no separate book-only limit. Going over is refused withtoo many subscriptions. Unsubscribing frees the slot immediately, and so does a book subscription that the server drops.
Resolved limits, not the table
Limits are resolved per account: every key on the account connects with the highest plan held by any of its active, entitled keys (with enterprise overrides applied). A Free key on an account that also has an entitled Pro key connects with Pro limits. Revoked and suspended keys cannot connect and do not count toward the account’s plan. Theconnected frame reports what your account may do, with enterprise overrides already applied. Enterprise numbers are negotiated per account and can sit above or below the Pro row, so read the frame rather than this table when it matters. The same resolved numbers back the meters on the portal’s Usage page.
Fair use
Bandwidth is measured as the payload bytes of data frames written to your connections, rolled up per hour, and shown against your allowance on the Usage page. It is a soft cap: nothing is cut off at the line, and the counts are at-least-once analytics rather than a billing record. The firehose is where it goes — roughly a hundred and thirty fills a second across every market — so an indexer left running ontrades is the case the allowance exists for.
An order-book snapshot is a data frame written to your connection, so its bytes count too: about 30 MB for every subscribe to l4Book on BTC (about 15 MB on HYPE, 11 MB on ETH, and about 5 KB for the median coin), then the per-block diffs. A client that resubscribes on every reconnect — which is the correct behaviour — pays the snapshot again each time. The TWAP status snapshot is a data frame on the same terms, though it carries one entry per live TWAP rather than a whole book.
When you hit a limit
See the wire protocol for the full error list and the overview for refusals.

