For stores and integrators

Service commitments

Version of 9 November 2026

Every month's availability, as measured: Service status

This document says what we commit to on availability and response times, how we measure it, and where anyone can check it. Every objective is below what we have already measured: we do not promise a number our systems have not already held. The measurement is the one behind the status page (/status, and the API's GET /v1/status, no key needed), which publishes every month, for every component, the availability computed with the rule of this document and the objective next to it.

1. In short

Component (id on the status page)What it isMonthly availabilityMaximum downtime in a 30-day month
assistantThe conversation that builds the cart and the order, on every channel99.9%43 minutes
web_chatThe chat on the shop's website99.9%43 minutes
apiThe REST API (/v1) for catalog, customers, orders, events, and the catalog feed99.9%43 minutes
eventsSigned webhooks, pull queue and order emails99.9%43 minutes
portalThe merchant portal99.9%43 minutes
mcpThe MCP server for AI agents99.9%43 minutes
Response time (server side, last 30 days)Median (p50)95th percentile (p95)
One assistant turn: from message received to reply ready≤ 9 s≤ 15 s
One API read (GET /v1/…)≤ 0.75 s≤ 2.5 s
One API write (PUT, POST, DELETE /v1/…)≤ 1 s≤ 3.5 s

2. How availability is measured

What is checked. Every minute, with no human involved, our alarms:

  • call the door of every component (a probe) and check that it answers as it should within 8 seconds;
  • check the real work of the last 5 minutes: server errors (5xx) of the functions and of the API, assistant turns that ended in error, webhook and email deliveries abandoned or late, scheduled jobs that stopped.

This produces the incidents, the same ones the status page shows:

  • outage: the component's door does not answer as it should to two probes in a row;
  • degraded: the door answers, but part of the work fails (one of the real-work checks is red).

An incident is closed after 10 minutes without the failure.

The monthly rule. For every component and every calendar month (UTC):

  • measured minutes = the minutes of the month in which the alarms measured the component. A minute without a measurement (alarms stopped, probe without an outcome) counts neither as available nor as down, and the page says so;
  • downtime minutes = the measured minutes covered by an outage or degraded incident of the component, from its first to its last red minute. Two overlapping incidents count once;
  • monthly availability = (measured minutes − downtime minutes) ÷ measured minutes × 100.

We also count degraded minutes, not only outages. It is the stricter choice for us: a failure that does not stop the door but stops your work (an order that does not go out, a turn that fails) is downtime even if the page answers.

Where to read it. GET /v1/status (no key) returns for every component availability.month (the current month, so far) and availability.previous_month (the closed month), with measured_minutes, outage_minutes, degraded_minutes, downtime_minutes, availability_percent, objective_percent and objective_met. The 30- and 90-day windows (availability.30d, availability.90d) are unchanged: their uptime_percent counts outages only.

What the measurement does not see. The probe proves that the door answers; the real-work checks see the errors that the systems record. A wrong answer that no system records as an error is not downtime for this measurement: our quality controls follow that, not this commitment.

3. Exclusions

Minutes due to the following are not downtime:

  1. maintenance announced on the status page at least 48 hours ahead, at most 4 hours a month, with its time;
  2. failures of the language-model providers (today OpenAI) recorded as such on the incident, for the assistant component only and only for the turns that depend on them;
  3. features declared as beta in the documentation, use beyond the limits of one's plan, failures of the customer's network or systems, account suspension, force majeure.

Today no exclusion is applied. The status page has no public maintenance register yet and no cause written on the incidents: until it does, the published calculation counts every minute of downtime, and is as strict as it can be. An exclusion will apply only once it is written on the status page together with the incident, so that anyone can recompute it.

4. Response times

Times are measured on the server side: from the arrival of the request to the reply being ready. They do not include the network up to you nor, for WhatsApp, the message's path through Meta.

  • Assistant: the total time of the turn recorded in the trace of every real turn (test turns and turns that ended in error excluded), over the last 30 days. A turn includes the call to the language model, which alone is the largest part.
  • API: the duration recorded in the register of every call (/v1, 5xx errors excluded), over the last 30 days.

These are objectives, not guarantees: they are published and measured, but they give no right to credits. Below 100 samples in the period a percentile is not judged: it reads «not measured».

5. What these numbers rest on (measured on 9 November 2026)

  • Availability: the alarms have been measuring since 24 October 2026 at 12:51 UTC (the MCP server since 5 November at 15:26). On 9 November, over about 22,600 measured minutes (15.7 days), no outage minute; 10 degraded minutes for assistant (two episodes of turns in error on 28 October) and 10 for portal (two episodes of 5xx errors on 31 October). Availability with the rule of this document: assistant and portal 99.956%, the others 100%. The month of October (from the 24th, 10,748 measured minutes): assistant and portal 99.907%. The history is short: that is why the objective is 99.9% and not higher.
  • Assistant turns: 575 real turns in 30 days (WhatsApp and web chat): median 6.9 s, p95 13.5 s, p99 18.2 s. Week by week the median was between 6.2 and 8.0 s and the p95 between 10.8 and 14.2 s.
  • API: 579 reads and 362 writes in 30 days: reads median 0.46 s and p95 2.15 s; writes median 0.63 s and p95 2.96 s. They are calls from our test benches (test environment): the API has no production traffic yet. The API objectives will be reviewed on the first month of real traffic.

6. How we communicate failures

  • The status page (/status) and GET /v1/status update themselves every minute: an incident appears as soon as the alarms open it, with the component, the severity, the start, the last minute seen and the resolution. Nobody has to write it by hand, so it cannot be kept quiet.
  • The status can be read by a program (JSON, ETag, 30-second Cache-Control): integrators can connect it to their own monitoring.
  • If the measurement itself stops, the page says so (unknown, and a 503 STATO_NON_MISURABILE when the status cannot be read): never «all good» by default.
  • The alarms notify our team of every incident they open.

7. Credits and contractual terms

This document provides no service credits. Different terms (credits, support response times, direct communications) are agreed in writing in the enterprise plan contract.

8. How they compare with the market (official pages read on 9 November 2026)

  • Twilio (twilio.com/en-us/legal/service-level-agreement/twilio-apis, updated 9 April 2026): 99.95% on the APIs, 99.99% for the Enterprise Edition; calendar month, minutes of unavailability; excludes maintenance announced 24 hours ahead, carrier failures and interruptions under 5 minutes; 10% credit.
  • Supabase (supabase.com/sla): 99.9% per calendar month, Enterprise plan only; excludes announced maintenance and cloud providers; credits from 10% to 30%.
  • Stripe (stripe.com/legal/ssa, 28 September 2026): no contractual availability commitment; publishes historical availability.
  • Shopify (help.shopify.com, Plus plan): «might include» a 99.99% in the individual contract, not public.

Our 99.9% is the level infrastructure providers publish for their enterprise plans, and it rests on what we have measured, not on a commitment someone else made for us.

9. How this document changes

Objectives go up when the measurement holds them for at least three complete calendar months, and the date is written. They do not go down because of a bad month: that month stays published with objective_met: false. Every version is dated.