Limits

These are the limits that belong to the platform rather than to your plan: how many tools an agent may have, how large an upload may be, how many of your generations may be unfinished at once, and more — and the defaults your agents run on where neither the agent nor your organisation sets one: the model, when a conversation is continued or greeted again, what a reply recalls of a person and when a conversation is read for facts, how much knowledge goes in the prompt, how a tool call is bounded. Your plan’s own limits — its requests an hour and at once, its voices, agents and end users — are on Usage.

The platform’s staff may change them, so this page lists them as they stood when these docs were built, and every page of the API reference names a limit by its key rather than by a number. A request past a limit is refused with the status the reference gives, and the error’s message names the number.

Limit	Key	Now	Built in
Requests a minute, when a plan names none	limits.org.requests_per_minute	300	300
Uploaded files in an organisation’s library	limits.org.asset_uploads	2,000	2,000
Voices an organisation may keep	limits.org.voices	25	25
Knowledge entries an organisation may hold	limits.org.knowledge_entries	1,000	1,000
Unfinished images an organisation may hold	limits.org.active_generations.image	8	8
Unfinished videos an organisation may hold	limits.org.active_generations.video	4	4
Unfinished queued speech generations an organisation may hold	limits.org.active_generations.tts	6	6
Unfinished queued transcriptions an organisation may hold	limits.org.active_generations.stt	8	8
Unfinished skyboxes an organisation may hold	limits.org.active_generations.skybox	2	2
Tools in an organisation’s library	limits.org.tools	200	200
Tools on one agent	limits.agent.tools	20	20
Conversations one end user’s key may open a minute	limits.end_user.opens_per_minute	60	60
The longest an ephemeral key may live	limits.ephemeral.max_ttl_seconds	30 days (2,592,000 seconds)	30 days (2,592,000 seconds)
The most credits one ephemeral key may hold	limits.ephemeral.max_credits	100,000 credits	100,000 credits
An uploaded image’s size	limits.uploads.image_bytes	40 MiB	40 MiB
An uploaded recording’s size	limits.uploads.audio_bytes	200 MiB	200 MiB
A voice’s recording’s size	limits.uploads.voice_bytes	20 MiB	20 MiB
The model of an agent that names none	agents.model	gemini-3.1-flash-lite	gemini-3.1-flash-lite
Quiet before a continued conversation is greeted again	agents.greeting.resume_after_minutes	15 minutes	15 minutes
How old a conversation may be and still be continued	agents.sessions.resume_max_age_days	7 days	7 days
Rounds of tool calls in one reply	agents.tools.max_rounds	3	3
Tool calls run at once	agents.tools.calls_per_round	5	5
How long an ask-first tool waits for a yes	agents.tools.confirm_for_minutes	10 minutes	10 minutes
A tool’s answer, at most	agents.tools.result_max_bytes	16,384 bytes	16,384 bytes
A conversation reports its tools’ failures every	agents.tools.report_every_minutes	5 minutes	5 minutes
What a reply recalls of a person, in characters	agents.memory.recall.budget_chars	6,000	6,000
Recall’s share for the most recent facts	agents.memory.recall.recent_share	0.25	0.25
What one reading reads of a conversation, in characters	agents.memory.reading.transcript_chars	200,000	200,000
Messages a reading reads before what is new	agents.memory.reading.context_messages	6	6
A conversation is read for facts after	agents.memory.facts.idle_minutes	15 minutes	15 minutes
A failed reading is tried again after	agents.memory.facts.retry_minutes	30 minutes	30 minutes
A replaced or removed fact is kept, hidden, for	agents.memory.facts.replaced_kept_days	30 days	30 days
A conversation is compacted at	agents.memory.compaction.default_threshold_tokens	32,000	32,000
Messages a compaction keeps word for word	agents.memory.compaction.keep_messages	6	6
A compaction’s notes, at most	agents.memory.compaction.max_output_tokens	4,000	4,000
Knowledge put in the prompt, in auto	agents.knowledge.inline_budget_tokens	8,000	8,000
Knowledge put in the prompt, at most	agents.knowledge.prompt_cap_tokens	32,000	32,000
Pieces a lookup hands the model	agents.knowledge.default_results	4	4
A lookup’s least similarity	agents.knowledge.min_score	0.6	0.6
A tool’s failures that mail its organisation	agents.tools.failures_to_mail	20	20
A tool’s failures are counted over	agents.tools.failure_window_minutes	60 minutes	60 minutes
A failing tool mails at most every	agents.tools.mail_every_hours	24 hours	24 hours
A streamed transcription’s pieces at ElevenLabs at once	speech.stt.stream_concurrency	6	6
An image not done this long after its submit fails	jobs.generation.deadline.image_minutes	15 minutes	15 minutes
A clip not done this long after its submit fails	jobs.generation.deadline.video_minutes	30 minutes	30 minutes
A skybox not made this long after its submit fails	jobs.generation.deadline.skybox_minutes	20 minutes	20 minutes
A skybox’s exports not ready this much later fail	jobs.generation.deadline.skybox_export_minutes	20 minutes	20 minutes
Requests’ contents are kept for	platform.retention.requests_days	90 days	90 days

Reading them from code

GET /v1/platform/limits answers them as they stand now. It needs no key: every limit is listed in one page, each with its key, label, value, unit and the platform’s built-in default. A size is in bytes, a time in its unit — ms, seconds, minutes or days — and credits are whole credits; a model is its id.

curl

curl https://api.lumavox.ai/v1/platform/limits

From your server, @lumavox/server reads them with limits() (Server SDK).