Features / MCP agent access
MCP session replay server for AI agents
Your coding agent already has the codebase. This is the endpoint that gives it the users too.
What it does
-
Eight tools, one endpoint
list_sites,list_tapes,read_transcript,get_stats,site_digest, and the alert tools — one stateless JSON-RPC endpoint, no per-agent glue code to write or maintain. -
Filtered queries, not a firehose
list_tapesfilters onhas_errors,has_rage_clicks,url_contains,min_duration_seconds, and date range, with a one-line summary per result — an agent picks the right session without reading forty transcripts first. - A read-scoped token, separate from ingest The MCP token is a distinct, revocable credential from the Agent access page. It cannot touch billing, delete a recording, or change a site — the only write surface it has is creating an alert.
- Works with what you already run Claude Code, Claude Desktop, Cursor, or a client you wrote yourself — Model Context Protocol is a standard, not an integration, so any compliant client discovers the same tools without custom wiring.
What it looks like
$ claude "users say coupons are broken — find it and fix it"
▸ usertapes list_tapes(has_errors: true)
← 14 tapes · rage clicks on "Apply coupon"
▸ usertapes read_transcript(3e7128f8)
← 0:19 click button "Apply coupon"
0:20 ⚠ POST /api/coupon · 422
0:23 ⚠ rage click ×4 · same button
▸ diagnosis: checkout.js swallows 422s — no error shown, button stays live
▸ editing checkout.js +14 −3
Every other tool in your agent’s reach — the repository, the test output, the git log — describes what the code is supposed to do. None of it describes what actually happened when a real person used it. That gap is why “a user says checkout is broken” still turns into an afternoon of clicking around a codebase looking for a plausible cause.
MCP closes it by making sessions queryable the same way your agent already queries everything else.
Why a protocol instead of an export
The alternative to MCP is exporting something — a CSV, a JSON dump, a pasted screenshot — and hoping it survives being pasted into a prompt. That works once. It does not scale to “ask this question every week,” and it puts a human back in the loop exactly where the point was to remove one.
A protocol means the server describes its own tools and any compliant client can call them, so the integration is written once, by us, and every agent gets it for free.
The tool that does the work
list_tapes is the one worth understanding, because it’s the difference
between an agent that reads forty transcripts to find the interesting session
and one that reads one. It takes has_errors, has_rage_clicks, url_contains
and a date range, and returns a one-line derived summary per tape — often
enough on its own for the agent to know which session to open.
read_transcript then returns that one session as markdown: every navigation,
click, and stall, with the moments worth attention marked. Timestamps line up
with the human player’s scrubber exactly, so when the agent cites “0:19” you
can go watch that second.
What the token can and cannot do
The MCP token is deliberately narrow. It reads sites, tapes, and transcripts, and it can create or delete a saved alert — a standing watch like “tell me when a session hits an error on /checkout.” It cannot delete a recording, change account settings, or touch billing. If it leaks, the blast radius is “someone can read your session data,” not “someone can do anything to your account.”
That’s a different credential from the site token in your tracking snippet, which is ingest-only and sits in public page source by design. Confusing the two — putting the MCP token somewhere client-side — is the one genuinely dangerous mistake available here, and it’s worth checking for explicitly the first time you wire this up.
What changes once it’s connected
Before MCP, “did anything break for users today” is a question you have to remember to ask, by opening a dashboard. After, it’s a question your agent can answer unprompted, correlate against the code it already has open, and — if you let it — fix. For the connection walkthrough, see session replay over MCP; for the endpoint, auth, and full tool schema, see the MCP server docs.