Ask your agent what your users ran into this week

You already have the agent. It cannot see your users.

Your agent reads the code, the diff and the failing test. But it never saw a person use what you built.

A finding is one kind of problem, such as a failed request on your checkout page. Sequential counts the people and sessions where it happened and keeps sample sessions as evidence.

Also, a model judges sessions against a set of questions. For each session that it judges, it says what went wrong.

One morning, start to finish

Example: Northlight Outfitters, our fictional demo shop.

1. 01

   You ask your agent

   "What is stopping people at checkout this week?"

2. 02

   Sequential answers

   38 people had a failed request at the shipping address. Three sessions show it. Each has a replay link and the second it happened. You open one: a postcode typed with a space, sent back.

3. 03

   Your agent acts, in Jira

   "File it for the web team." It creates WEB-231 with those sessions, with its own access. Sequential has no Jira connection.

4. 04

   Your team ships

   Your team ships release 2026.10.09. Your CI tells Sequential which finding the release is for.

5. 05

   Sequential measures

   41% → 12% of sessions on the shipping page: the share of sessions where the problem showed up.

Sequential reports what changed. The cause is yours to judge.

Set up in three steps

1. 01

   Put Sequential on your site

   Create an account, and the Setup page gives you an install prompt for your coding agent.

2. 02

   Connect your agent

   Make a token with the recommended option, "Agent: read everything, plus Ask". It is for one project. It includes session search and it cannot write. Then paste the ready-made command. The token is shown once: treat it like a password.

3. 03

   Ask a first question

   Try: "Using Sequential, what should I know about my product this week?"

claude mcp add --transport http --scope user sequential <api>/mcp \
  --header "Authorization: Bearer <token>"

Scripts and CI get the same tools and answers over REST, with the same token.

POST <api>/api/agent/<tool>

What your agent can ask for

Your agent sees only the tools that its token allows.

• What is wrong. Ranked problems, each with a fix brief: what users went through, where, since when, and the evidence.
• The sessions. What went wrong in each session, with key moments linked into the replay.
• Plain-words search. Up to 50 sessions for each call.
• Your numbers. Events, funnels, flows, retention, users and accounts, with people as ids.
• Before and after. The measurement around a release.

Built for a reader that you do not control

Some text that your visitors type can try to trick an agent. Sequential checks that text first.

• Redacted. Emails, tokens and card numbers come out.
• Screened. A model removes passages that read like an instruction to an agent.
• Held back if unchecked. Text that Sequential cannot check is not sent. If a fix brief cannot be fully checked, Sequential refuses it.
• Labelled. Visitor text is marked as data, not instructions.

You stay in charge

• Read-only unless you say so. A tool that writes needs a permission that you select. The recommended option has none of them.
• One project for each token. A token expires after 30, 90 or 365 days. Only owners and admins create or revoke tokens. The Agents page shows what each token called.
• Spend with a limit. Only session search spends, from its own daily allowance for each project and for each token. If the allowance is used up, or Sequential cannot read it, the search is refused before it spends. If a limit stops a search part-way, the answer says so.
• No spend at all. Select "Read everything, no Ask" (no session search), and your agent cannot spend from the allowance.

Good to know

• Ids arrive as your site sends them. Your agent never gets names, emails or IP addresses. But if your user ids are emails, your agent sees emails.
• Screening is a filter, not a guarantee. It can miss an instruction that is split across rows.
• Up to 30 days for each question. The hosted preview keeps 7 days of data by default.
• Session judgements come from a model, and they can be wrong. Your agent can see how often they were right for each question.
• A before-and-after needs a recorded release that names the finding. It reports what changed, never why.

Questions