Operational limits

The caps emisar enforces so one noisy action, one busy agent, or one aggressive export cannot overwhelm a page or the API. Most are generous. These are the ones worth knowing before you build against them.

Each row shows what you will see at the cap; the name links to the page that owns the behavior.

What	Cap	At the cap
Action output	Set per action in the pack, separately for stdout and stderr — 64 bytes to 16 MiB across the shipped catalog	Bytes past a stream's cap are dropped from the delivered output but counted, and stdout and stderr are each flagged truncated on their own.
Live output in the console	The most recent 500 progress lines	Older lines are trimmed from the browser view with a note. Run history keeps accepted chunks within a durable budget — up to 50,000 chunks and 64 MiB per run, one chunk capped at 256 KiB. The host journal keeps bounded previews, byte counts, and hashes of the complete redacted streams.
MCP list pages	15 per page by default, with a maximum of 50. `recent_runs` supports up to 100. `find_actions` supports up to 15.	Continue with the cursor the previous response returned. A catalog cursor is bound to its exact query and expires after 15 minutes.
MCP long-poll wait	Up to 60 seconds per call	`run_action` and `wait_for_run` return the current state at the deadline. One credential can hold 8 concurrent waits per node. Every other tool answers immediately.
MCP structured output	8 KiB per run, shared down when one response carries several	A run whose parsed output is larger is dropped from the response and replaced with a `wait_for_run` continuation that fetches it on its own.
MCP response size	512 KiB per tool result	Tools shrink themselves first — fewer rows, shorter output tails. A result that still does not fit returns a transport error carrying a `get_operation` continuation.
Runbooks	64 KiB definition. 32 inputs, 16 stages, 16 target refs per step, and 256 resolved items. A wait polls every 5 seconds to 1 hour for 2 to 100 attempts. A parallel stage runs 1 through 16 attempts at once, 5 by default. A title is at most 80 characters and 320 bytes, a description 4,096 characters and 8 KiB.	Authoring or preflight rejects the definition before an execution is created. The MCP read budget is derived from these limits, so a saved runbook always fits one response.
Runbook execution	1 MiB frozen plan. 24 hours end to end, waits included	Preflight rejects a plan that resolves larger. An execution still running at the deadline halts as timed out.
Audit export	100 events per page, 1,000 maximum	A full page carries `Link: rel="next"` and means more may be waiting — keep paging from `X-Next-Cursor` until it disappears.
Audit export requests	60 requests a minute per token	Further requests answer `429`. A poller must hold its cursor and retry rather than treat the failure as an empty page.

Scripting against the API?

The fixed tools, identifiers, and recovery semantics of the emisar MCP server — including how a retried dispatch resolves to its original operation — are in the MCP CLI & reference.

← Previous

Security model