Session replay MCP server: run UserProof from Claude Code, Codex or Cursor
UserProof is a session replay MCP server with 14 tools, built for founders who build with a coding agent. Install the plugin in Claude Code, or add one stdio server in Codex, Cursor or any MCP agent. The agent sets up recording, then returns the top fix from real visitor sessions. Free covers 3 sites and 5,000 visits a month.
We counted the tools in the server source on November 8, 2026: 14 are registered. A live tools/list run on November 7, 2026 also listed 14 (output on the install page). PostHog's tools reference, read the same day, lists 9 for session replays alone, inside a server that covers every PostHog product. UserProof's 14 do one job: find the fix, then check it worked.
Jump to: which agent, which install · the 14 tools · set it up · limits · where others win · price · FAQ
Which agent, which install?
Every agent below runs the same server. Claude Code gets it as a plugin; the others start it as a local stdio MCP server with npx.
| Agent | How it connects | Command or file | Steps |
|---|---|---|---|
| Claude Code | Plugin (includes the MCP server and the setup and review commands) | claude plugin marketplace add GetUserProof/userproof && claude plugin install userproof@userproof | Set up in Claude Code |
| Codex (CLI or IDE extension) | Local stdio MCP server; raise the start timeout to 60 seconds | codex mcp add userproof -- npx -y github:GetUserProof/userproof | Install in Codex or Cursor |
| Cursor | Local stdio MCP server | An entry named userproof in ~/.cursor/mcp.json | Install in Codex or Cursor |
| ChatGPT desktop app | Local stdio MCP server, shared with Codex | Settings, MCP servers, same npx command | Connect ChatGPT |
| VS Code, Windsurf, Claude Desktop or another MCP agent | Local stdio MCP server | npx -y github:GetUserProof/userproof, written in each agent's own config format | MCP config generator |
| ChatGPT in a web browser | Not supported: no remote server | None | Why not |
What can the agent do with the 14 tools?
It can read the top fix, set up and prove recording, and check a shipped fix. We counted the tools in the server source on November 8, 2026 (14 registered; server version 3.10.0, as the server reports itself) and read what each does there. Every tool answers in plain text, so any agent can use it.
| Job | Tool | What it returns |
|---|---|---|
| Read findings | get_top_fix | The single highest-impact fix on the site, the sessions that prove it, and dollar math when a conversion event is set. It adds a notice when the plan's visit limit is near. |
fleet | All your sites in one read: the top fix or biggest leak for each, plus patterns that repeat across sites. | |
| Pick the site | list_sites, select_site | The sites on the account with their ids, and the choice of which one this project uses. The agent saves it to .userproof in your repo. |
| Set up recording | wait_for_first_event | Waits up to 90 seconds for the first event from your site and says what arrived and from where. If nothing arrives, it returns the likely cause. |
check_coverage | Which pages in your sitemap are missing the script tag. | |
set_goal, set_industry | The goal (a click or a page visit, with an optional dollar value) that findings are ranked by, and an industry tag for anonymous peer comparisons. | |
| Before traffic | scan_site | A read-only scan of the live site that returns the one worst defect it can reproduce 3 of 3 times. Takes 1 to 3 minutes; 10 a day per site. |
| Close the loop | mark_fix_applied, check_fix | Records that a fix is live, then compares the finding's signal, the share of sessions that go on to a next step, and conversion before and after. It says "inconclusive" when traffic is thin. |
| Account | get_usage, get_activation, upgrade | Visits used against the plan limit, the minutes from signup to each setup milestone, and a checkout link for Pro or Scale. |
The Claude Code plugin adds commands on top of the tools: /userproof:setup and /userproof:review. The review command gets the top fix, finds the code to change and, after you ship, runs the check. For what counts as a dead or rage click, see how UserProof finds them.
How do I set it up?
You do four things; the help pages hold the exact steps so this page does not repeat them.
- Connect the server to your agent. Use the row for your agent in the table above. Done looks like the agent listing
userproofas an enabled MCP server (codex mcp listorclaude plugin list), then a restart so the tools load. - Ask the agent to set up UserProof. In Claude Code that is
/userproof:setup; elsewhere you tell it to readhttps://getuserproof.com/setup.mdand follow it. You open one sign-in link in your browser. Done looks like the agent reporting statusreceived: the script tag records. - Let visitors arrive, then ask "what should I work on?" Done looks like one problem with the sessions that prove it. The count and thresholds are in the dead and rage click rules.
- Ship the fix and mark it. Done looks like the fix recorded as applied; the check then compares before and after.
The key is written to ~/.config/userproof/key by the sign-in step, so nothing secret goes into a config file. The plugin reads only .userproof in your repo and that key, and does not upload your source code.
What can the agent not do?
- It does not change your code by itself. No tool edits files. The server returns the finding and your agent makes the change, if you tell it to.
- It does not list or play raw sessions. Sessions come attached to a finding. There is no tool to browse every recording, and none to delete one.
- It runs locally over stdio only. There is no remote MCP server URL, so web-based chat clients that need one cannot connect.
- Thin traffic gives thin answers. A finding needs enough sessions, and
check_fixsays "inconclusive" until there is enough traffic. A funnel read when nothing passes the bar is a lead, not proof. - Free keeps recordings for 7 days. Pro and Scale keep 90 days.
- No heatmaps and no surveys. Use a tool from our ranking of replay tools if you need those.
- It needs Node 18 or newer for the stdio server.
Where do PostHog's and Clarity's MCP servers still win?
They win when you already run that product and want its data, which ours does not hold. Both facts below were read on November 8, 2026.
- PostHog runs a free hosted server at
mcp.posthog.com/mcpthat works with Claude Code, Cursor, Codex, VS Code, Windsurf and Zed. Its tools reference lists 9 session replay tools (including recording lists, console logs, playlists and deletes) and more for feature flags, experiments, error tracking, logs, surveys and SQL. Ours has no SQL, flags, experiments or deletes. Some PostHog tools use LLMs and may bill as PostHog AI spend. Side by side: PostHog MCP vs UserProof. - Microsoft Clarity has an MCP server with 3 tools: a plain-language dashboard query, a filtered list of session recordings and a documentation lookup. Clarity itself is free, records up to 100,000 sessions per project per day and has heatmaps. Its Data Export API allows 10 requests per project per day, 3 dimensions per request and 1,000 rows per response. Ours cannot filter recordings by browser or country. Setup is in Clarity MCP in Claude Code.
The difference is in what comes back. Those servers return your data for you to question. UserProof's returns one ranked fix and then checks it. If you want to explore, pick theirs; if you want to be told what to change, pick ours.
What does it cost?
| Plan | Price | Recorded visits a month | Sites | Recordings kept |
|---|---|---|---|---|
| Free | $0 | 5,000 | 3 | 7 days |
| Pro | $29 a month | 50,000 | Unlimited | 90 days |
| Scale | $99 a month | 250,000 | Unlimited | 90 days |
Read from our pricing page on November 8, 2026. Every tool is on every plan. Past your plan's visits, UserProof keeps recording 1 in 4 visitors; nothing is cut off. To write the config for a second agent, use the MCP config generator, which runs in your browser.
To see this server beside the others a founder might add, read MCP servers for coding agents, ranked.
Sources, read November 8, 2026
- UserProof MCP server source (14 registered tools, each tool's behavior), read November 8, 2026; live
tools/listoutput from November 7, 2026 on the install page - UserProof setup.md (install commands, key location, stdio server), read November 8, 2026
- UserProof GitHub README (plugin and Codex and Cursor install, what data is collected), read November 8, 2026
- UserProof plans: pricing page, read November 8, 2026
- PostHog, Model Context Protocol docs (hosted free server, supported clients, AI spend), read November 8, 2026
- PostHog, MCP tools reference (9 session replay tools), read November 8, 2026
- Microsoft, Clarity MCP server README (3 tools), read November 8, 2026
- Microsoft Learn, Clarity Data Export API (10 requests per project per day, 3 dimensions, 1,000 rows), read November 8, 2026
- Microsoft Learn, Clarity FAQ (100,000 sessions per project per day), read November 8, 2026
FAQ
What is a session replay MCP server?
It is a program your coding agent starts that answers questions about recorded visitor sessions. UserProof's has 14 tools. The main one, get_top_fix, returns the highest-impact problem with the sessions that prove it. PostHog and Microsoft Clarity also publish MCP servers that read their own replay data.
Which coding agents can run the UserProof MCP server?
Claude Code through its plugin, and Codex, Cursor, the ChatGPT desktop app, VS Code, Windsurf or Claude Desktop through a local stdio server started with npx. It needs Node 18 or newer. ChatGPT in a web browser cannot use it, because there is no remote server.
Is an MCP session the same as a session replay session?
No. An MCP session is the connection between an agent and a server. A replay session is one visitor's visit to your site. UserProof's server runs over stdio on your machine, so there is no remote connection to keep alive. The sessions it reads are your visitors' recorded visits.
Does the agent edit my code or watch recordings by itself?
No tool edits code or plays a video. The server returns text: the finding, the sessions behind it and the numbers. Your agent decides what to change in your repo, and in Claude Code the review command finds the code for you. Nothing is marked fixed until you say the change is live.
Does running UserProof from an agent cost extra?
No. The server is free to run, and analyses and fix checks are unlimited on every plan. Plans are priced on recorded visits: Free is $0 for 3 sites and 5,000 visits a month, Pro is $29 a month for 50,000, Scale is $99 a month for 250,000.