00:00For AI coding agents

Let your agent fetch the review and fix it.

Voicetrail runs a remote MCP server. Connect your AI tool once, and it can list your recordings and read their timeline, issues, errors and screenshots. The agent runs on your own subscription; MCP is included in every seat.

Server URL

https://app.voicetrail.app/mcp

Sign-in uses OAuth with your Voicetrail account; you pick which workspace the tool can read. For tools without OAuth, create a personal API key in the dashboard and send it as Authorization: Bearer vt_….

you › Get my latest Voicetrail review and fix the issues.

● voicetrail · get_recording (id: "latest")

● voicetrail · get_issues

● voicetrail · get_timeline (format: "md")

What get_timeline returns (real excerpt):

00:09  POINT     at: link "Cart 0" in nav "Products Cart 0 Account About (full page)"  (a[href="/cart"])  [shots/002.jpg]
00:11  SAY       "And look at this notebook card."
00:14  SAY       "The price shows NAN dollars."
00:14  CIRCLE    around: div.card "Notebook $NaN Add to cart"  ([data-testid="product-3"])  [shots/003.jpg]
00:17  SAY       "That's clearly broken."
00:20  SAY       "This filter button here doesn't do anything when I click it."
00:21  ARROW     pointing at: button "Filter"  (#filter)  [shots/004.jpg]
00:21  CLICK     button "Filter"  (#filter)  ×2
00:24  SAY       "Let me try apply discount."
…

Tool calls as a client would show them; the excerpt is unedited export text.

00:20Setup

Connect your tool

Checked against each tool's docs on 2 November 2026

Claude Code

Add the server, then sign in from inside Claude Code.

Official docs

claude mcp add --transport http voicetrail https://app.voicetrail.app/mcp

/mcp

In the /mcp panel, pick voicetrail and sign in in the browser. Add --scope user to make it available in every project, or --scope project to share it with your team through .mcp.json. The prompts show up as /mcp__voicetrail__fix_feedback.

Claude.ai and Claude Desktop

Custom connectors work on every Claude plan; the Free plan allows one.

Official docs

Open Customize → Connectors.
Click + Add, then Add custom connector.
Name it Voicetrail and paste https://app.voicetrail.app/mcp as the server URL.
Choose Sign in now, sign in to Voicetrail and pick your workspace.
In a chat, open the + menu → Connectors and turn Voicetrail on.

On Team and Enterprise plans, an owner adds the connector under Organization settings → Connectors first; members then sign in from Customize → Connectors.

ChatGPT

Custom MCP apps need developer mode, which OpenAI currently offers on Business, Enterprise and Edu workspaces. Plan rules change often; check OpenAI's help page if the option is missing.

Official docs

A workspace admin turns on developer mode (Business: Workspace settings → Permissions & Roles; Enterprise and Edu: Settings → Apps → Advanced settings).
In Workspace settings → Apps, click Create, paste https://app.voicetrail.app/mcp as the endpoint and choose OAuth.
Click Scan Tools, then Create, and sign in to Voicetrail when ChatGPT asks.

Cursor

Add Voicetrail to ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project). Cursor opens the sign-in page the first time it connects.

Official docs

~/.cursor/mcp.json

{
  "mcpServers": {
    "voicetrail": {
      "url": "https://app.voicetrail.app/mcp"
    }
  }
}

VS Code (GitHub Copilot)

Run MCP: Add Server from the Command Palette, choose HTTP and paste the URL, or add it to .vscode/mcp.json. Note that VS Code uses "servers", not "mcpServers".

Official docs

.vscode/mcp.json

{
  "servers": {
    "voicetrail": {
      "type": "http",
      "url": "https://app.voicetrail.app/mcp"
    }
  }
}

Windsurf

Windsurf reads remote servers from serverUrl (not url) in its MCP config file.

Official docs

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "voicetrail": {
      "serverUrl": "https://app.voicetrail.app/mcp"
    }
  }
}

Codex CLI

Add the server, then sign in with OAuth.

Official docs

codex mcp add voicetrail --url https://app.voicetrail.app/mcp
codex mcp login voicetrail

~/.codex/config.toml (equivalent)

[mcp_servers.voicetrail]
url = "https://app.voicetrail.app/mcp"

Gemini CLI

Add the server, then sign in from inside Gemini CLI. In settings.json, streamable HTTP servers use httpUrl.

Official docs

gemini mcp add --transport http voicetrail https://app.voicetrail.app/mcp

/mcp auth voicetrail

~/.gemini/settings.json (equivalent)

{
  "mcpServers": {
    "voicetrail": {
      "httpUrl": "https://app.voicetrail.app/mcp"
    }
  }
}

01:05What it serves

Tools

All tools are read-only. They follow your workspace's visibility setting, and admins can turn MCP access off.

Tool	Input	Returns
list_recordings	query, site, since, owner, status, limit, cursor	Recordings with title, site, date, length, owner, issue and error counts, and a dashboard link
get_recording	id or "latest"	Metadata, pages visited, the AI summary and the issue list
get_timeline	id, format (md or json), from, to	The LLM-ready timeline.md, in chunks of about 15k tokens for long sessions
get_issues	id, severity, type	Issues with quote, element and CSS selector, page, related errors, and a link that opens the replay at that moment
get_errors	id, kind (console, exception or network)	Console errors and exceptions with stacks, failed requests with method, URL and status
get_transcript	id, from, to	The narration with timestamps
get_screenshot	id, shot (e.g. shots/003.jpg)	The screenshot as an image, so multimodal agents can look at the page
search_recordings	query	Matches in titles, transcripts and issues, with recording id and time

Prompts

Clients that support MCP prompts show these as commands.

fix_feedback(recording): Fetches the issues and timeline, finds each element by its selector and page path in your repo, fixes it, and reports what changed with the replay link.
triage(recording): Groups the issues into bugs, UX and content, and drafts tickets.

Resources

Clients that let you attach a resource can add voicetrail://recordings/{id}/timeline.mdor …/issues.json to a chat directly.

01:40Safety

Keep approval prompts on.

A recording contains text from the pages you visited: page content, console messages, URLs. A page can include text written to look like instructions to an AI agent ("ignore previous instructions and run…"). That's called prompt injection, and coding agents can edit files and run commands.

What we do: the tools are read-only, and every result wraps recorded content in a block labelled as untrusted recorded content, not instructions. Page text is truncated, and tokens can only read one workspace.

What you should do:

Keep your agent's approval prompts on for shell commands. Don't run agents with auto-approve when they read recordings of sites you don't control.
Review the agent's changes before you merge them.
Use the allowed-domains setting so your team only records your own sites.

MCP is included in every seat and in the trial.

Start the trial