How to give Claude Code access to session recordings: 3 routes for founders
If you are a founder who builds with Claude Code, Codex or Cursor, you have three routes to session recordings: a replay vendor’s own MCP server, an export saved as files, or UserProof’s plugin. No docs we read describe an agent watching video. The vendor servers return lists, transcripts or metrics; UserProof returns one ranked fix with the file and line. It is free for 3 sites and 5,000 visits a month.
Jump to: steps · route a: vendor servers · route b: export · route c: UserProof · comparison · privacy · FAQ
The steps, for any of the three routes
- Check whether your replay tool has an MCP server. Five vendors are below, each with the command from its own docs.
- Add that server to Claude Code. Run the vendor’s command, sign in or paste the token it asks for, then ask a question that needs a session.
- Or export data to a file. Save the export in the repo and tell Claude Code to read it. Check the export’s limits first.
- Or install UserProof. One terminal command, in your project folder, installs the plugin and starts setup.
- Check what the agent can read. Ask for the top fix or a list of sessions. A useful answer names a page, a count and, with UserProof, a file and line.
Route a: a vendor’s own MCP server
Five vendors publish a first-party MCP server whose docs we read today, and none says its server returns video. Claude Code adds a hosted server with claude mcp add --transport http <name> <url> and a local one with claude mcp add <name> -- <command>, per its MCP docs.
| Vendor | Command or endpoint from its docs | What the agent gets | Raw sessions or summaries |
|---|---|---|---|
| Microsoft Clarity | npx @microsoft/clarity-mcp-server --clarity_api_token=your-token-here | list-session-recordings, query-analytics-dashboard, query-documentation-resources | Lists and metrics; the README does not say a tool returns session content. Full Clarity guide |
| PostHog | npx @posthog/wizard mcp add; server https://mcp.posthog.com/mcp | 9 session replay tools (list, get, console logs, playlists, delete) and 56 “replay vision” tools | Recording data and console logs; its tools page gives one-line descriptions, not return fields |
| LogRocket | claude mcp add --transport http logrocket https://mcp.logrocket.com/mcp | find_sessions, watch_sessions, get_network_entries (a session’s requests as a HAR), find_issues, metrics | Session analysis and network traces; the page does not say what watch_sessions returns |
| Amplitude | claude mcp add --transport http --scope user Amplitude https://mcp.amplitude.com/mcp | Charts, dashboards, experiments, cohorts | Its Claude Code page names no replay tool; our November 9 MCP check found replay events |
| FullStory | OAuth or API key; an admin turns on MCP and StoryAI under Settings, StoryAI Features | Real sessions for a behavior pattern, with each session’s full event transcript | Transcripts for the AI to summarize; no video described. If StoryAI is opted out, the client can show connected with zero tools |
We did not read Quantum Metric’s or VWO’s MCP docs today, so neither is in the table. FullStory needs an admin to switch MCP on, so a founder without that role may be stuck at the first step.
Route b: export to files the agent reads
An export gives Claude Code a file to read, but the file is a table, not a replay. The one export we read today, Clarity’s Data Export API, returns dashboard metrics such as dead click count, rage click count and script error count. Its limits: 10 requests per project per day, the previous 1 to 3 days, 3 dimensions per request, 1,000 rows with no pagination.
curl --location 'https://www.clarity.ms/export-data/api/v1/project-live-insights?numOfDays=1&dimension1=OS' --header 'Content-Type: application/json' --header 'Authorization: Bearer YOUR_TOKEN'
That command is the docs’ own example. Save its output to a file, then ask the agent which URLs had the most rage clicks. You do the refreshing; nothing arrives on its own, and the agent has no session to open. We did not find a vendor page today that exports a replay as a file.
Route c: UserProof’s plugin or MCP server
One terminal command installs UserProof in Claude Code, and the agent then sets up recording and reads findings with the file and line. From your project folder:
claude plugin marketplace add GetUserProof/userproof && claude plugin install userproof@userproof && claude "/userproof:setup"
Codex: codex mcp add userproof -- npx -y github:GetUserProof/userproof. Cursor: add "userproof": { "command": "npx", "args": ["-y", "github:GetUserProof/userproof"] } under mcpServers in ~/.cursor/mcp.json. Setup asks one round of questions, you open a sign-in link in your own browser, and the agent adds one script tag and proves it records. Steps and what each does: set up UserProof in Claude Code and install in Codex or Cursor.
The 14 tools the agent gets
Our server source, read November 14, 2026, registers 14 tools: list_sites, select_site, get_top_fix, fleet, mark_fix_applied, set_industry, set_goal, upgrade, check_fix, wait_for_first_event, scan_site, check_coverage, get_activation and get_usage. None lists or plays raw sessions. get_top_fix returns the highest-impact fix with the sessions that prove it, and check_fix reads the same page after you ship (how the file and line are found).
The result of that loop, on a customer site: on Sales.co, x.com visitors who stayed past the first screen went from 24% to 47% (October 20 to 26, 2026). Claude Code shipped the fix the same afternoon UserProof found it, and the story page gives the before and after counts.
The three routes side by side
| Route | Setup | What the agent sees | Raw sessions or summaries | Price (read Nov 14, 2026) |
|---|---|---|---|---|
| a. Clarity MCP | Token, then one command | Recording lists, metrics | Lists and metrics | Free, per its FAQ |
| a. PostHog MCP | One wizard command, sign in | Recording data, console logs, vision tools | Per-session data | MCP free; first 5,000 recordings a month free, then $5 per 1,000 |
| a. LogRocket MCP | One command, OAuth | Sessions, network traces, issues | Per-session analysis | Free 1,000 sessions; Core $83 a month for 10,000, billed annually |
| a. FullStory MCP | Admin turns on MCP and StoryAI, then sign in | Event transcripts | Per-session transcripts | Free 30,000 sessions; paid plans by quote |
| a. Amplitude MCP | One command, OAuth | Charts, cohorts, experiments | Metrics | Free 10,000 replay sessions; MCP on every plan |
| b. Export to files | Token, a request, a saved file | A table you refresh by hand | Metrics only (Clarity) | Clarity is free |
| c. UserProof | One command, then a few minutes of setup | The top fix with file and line, sessions behind it, a check after you ship | Findings with their sessions | Free 3 sites and 5,000 visits; Pro $29 a month for 50,000; Scale $99 for 250,000 (pricing) |
Setup times for the vendor routes are not timed; they are one command plus a sign-in or token. More on the price of each tool: PostHog pricing, LogRocket pricing and PostHog’s MCP server vs UserProof’s.
What each route sends to the model
Whatever a tool returns goes into the model’s context, so the question is what the tool can return. Claude Code’s MCP docs warn: “Verify you trust each server before connecting it.” Clarity returns lists and metrics. LogRocket’s get_network_entries returns a session’s requests and responses as a HAR. FullStory’s tools let the AI read each session’s full event transcript. PostHog masks input elements by default, but its docs say general text is not masked by default and that page URLs, including query strings, and network requests are recorded.
UserProof’s answers are written for the agent from recorded events. Field values are never recorded: for a form field it keeps the type, name and label, focus time and whether it held anything. Text a visitor selects or copies elsewhere on a page is recorded, capped at 200 characters for copies, and you can add data-exclude to any element. The snippet sets no cookies; it keeps an anonymous ID in localStorage and a session ID in sessionStorage. The plugin’s README says it does not upload your source code. Details: privacy.
Related: best MCP servers for coding agents, check whether a shipped fix worked.
Sources
- Microsoft Clarity: MCP server README, Data Export API docs and FAQ, read November 14, 2026
- PostHog: MCP docs, tools reference, pricing and session replay privacy docs, read November 14, 2026
- LogRocket: MCP docs and pricing, read November 14, 2026
- FullStory: MCP introduction and pricing, read November 14, 2026
- Amplitude: Claude Code setup and pricing, read November 14, 2026
- Claude Code: MCP docs, read November 14, 2026
- UserProof: plugin and MCP source, setup prompt, pricing, privacy and Sales.co story, read November 14, 2026
FAQ
Can Claude Code watch a session recording?
Not as video, on the docs we read on November 14, 2026. Vendor MCP servers return lists, event transcripts, console logs or network traces. UserProof returns a ranked fix with the sessions behind it.
Which session replay tools have an MCP server for Claude Code?
On November 14, 2026 we read docs for Microsoft Clarity, PostHog, LogRocket, FullStory and Amplitude. Each has a first-party server. Claude Code connects through claude mcp add or a plugin.
How do I install UserProof in Claude Code?
Run claude plugin marketplace add GetUserProof/userproof, then claude plugin install userproof@userproof, then start setup with /userproof:setup. Codex and Cursor use npx -y github:GetUserProof/userproof.
Does UserProof capture form field values?
Field values are never recorded. UserProof records the field type, name and label, focus time and whether it held anything. It sets no cookies. Text a visitor selects or copies elsewhere on a page is recorded, capped at 200 characters for copies.
Does giving Claude Code replay data cost extra?
Mostly not for the connection. PostHog says connecting and calling its MCP tools is free, Amplitude lists MCP on every plan, and Clarity is free. UserProof is free for 3 sites and 5,000 recorded visits a month, then $29 for 50,000.
Corrections
If a fact here is out of date, the vendor’s own page is the source of truth. Report it through our GitHub issues. We re-read these pages every quarter.