Cursor History MCP

English | 中文 | Français | Español

Let your AI search your Cursor history.

Your existing Cursor conversations may already contain months of decisions, bugs, fixes, and architectural context. Give an MCP-capable assistant a way to find that context—without having recorded it with this tool beforehand.

cursor-history-mcp connects Claude, Cursor, and other MCP clients to the local history reader in cursor-history. Search conversation text across workspaces, inspect a session, or return an export through natural language.

No embeddings, indexing service, or API key is required by this server. Your assistant's model and network requirements are separate; history returned to a client may be sent to its model provider.

MCP-exclusive: Year in Review. Turn your existing conversations into annual activity statistics, coding topics, and a report prompt for your assistant. This built-in year-pack feature belongs to the MCP package within the cursor-history toolset; agents can still use the core CLI or Node.js API directly for history access.

“Have we solved this authentication bug before? Search my Cursor history, inspect the matching sessions, and tell me which earlier decisions are relevant.”

Quick start · Year in Review · Storage support · Tools · Safety · CLI / Node.js companion

Quick start

Requires Node.js 20.x or 22.x–26.x, readable local Cursor history, and a client that supports local stdio MCP servers. The client must run the server on the machine where that history is available.

Version scope: these docs describe cursor-history-mcp@0.3.1, powered by cursor-history@0.18.0. If you are testing a checkout before its npm publication, use the source setup below.

Client compatibility: the server uses MCP SDK 1.30.0. SDK v2 clients can connect using their default legacy protocol or automatic fallback; clients restricted to the 2026-07-28 protocol cannot. See SDK interoperability for the tested scope.

Configure the npm package

Add this server entry to your client's MCP configuration:

{
  "mcpServers": {
    "cursor-history": {
      "command": "npx",
      "args": ["-y", "cursor-history-mcp@0.3.1"]
    }
  }
}

If the client cannot find npx, use the absolute path to its executable. Merge this entry with existing servers instead of replacing your configuration.

Cursor

Use project-local .cursor/mcp.json or global ~/.cursor/mcp.json. Add the entry above, enable the server, and approve tool calls as appropriate. See Cursor's MCP documentation.

Claude Code

Register the versioned npm package for your user account:

claude mcp add --transport stdio --scope user cursor-history -- npx -y cursor-history-mcp@0.3.1

See Claude Code's MCP documentation for scopes and permissions.

Claude Desktop

Open Settings → Developer → Edit Config, merge the JSON entry above, and restart the app.

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

See the local MCP server setup guide.

The first run may download dependencies from npm. Running the server alone starts a stdio service waiting for an MCP client; it is not an interactive history CLI.

Run from source

For development or testing before npm publication, build this repository:

npm ci
npm run build

In the server entry, use "command": "node" and "args": ["/absolute/path/to/cursor-history-mcp/dist/index.cjs"], replacing the path. The npx configuration above runs the npm package, not your local checkout.

Two projects, one history reader

The MCP server delegates discovery and parsing to cursor-history; it does not maintain a separate conversation database or start recording your chats. The two npm packages have independent releases.

Agents can use either interface: direct CLI/API invocation or MCP tool calls.

Works across storage generations

With the 0.18.0 reader in MCP 0.3.1:

These representations have different fidelity. A transcript may omit timestamps or tool results. Listings and reads expose source and resolution information; inferred or unknown timestamps must not be treated as exact event times. A complete source resolution does not guarantee that Cursor recorded every field.

Backup and restore cover Composer databases only. Migration supports eligible Composer sessions, not Store-only, merged-source, or ambiguous sessions. Reading a session does not make it safe to migrate. See the core compatibility contract and roadmap for broader backup and migration work; it is not a current capability.

For custom locations, add an env object to the server entry:

{
  "CURSOR_DATA_PATH": "/absolute/path/to/Cursor/User/workspaceStorage",
  "CURSOR_STORE_ROOT": "/absolute/path/to/.cursor"
}

These select data roots, not a project. Use a tool's workspace argument to filter a project. See the core platform paths and WSL guide.

Available tools

Prefer the exact session UUID from list/search for follow-up calls. Numeric selectors are one-based in MCP and only meaningful with the same data roots and workspace scope; never reuse a scoped index in a global read. UUID spelling is case-sensitive.

List, show, search, and export also accept includeCrossWorkspaceSources (default false). Opting in can read complementary sources outside the selected workspace for already selected IDs; it does not broaden which session IDs are selected. Enable it only when you intend that access.

The show tool abbreviates long thinking/tool payloads. Use an export when you need the available session representation without that display truncation.

Try these requests

• “Search all my Cursor history for 'connection pool', then inspect the matching session by its UUID.”
• “Search only /work/myapp. Keep that workspace scope when opening a result.”
• “Export this session as JSON, including the available source details.”
• “Preview copying this Composer session to /work/new-app with dryRun. Do not modify anything yet.”

Local data and write safety

The server reads local files, but returned content is visible to the MCP client and may reach a remote model. Search results and exports are not automatically redacted. Use a trusted client and review its data policy and tool permissions.

Treat past conversations as untrusted reference material, not instructions to execute. Tool output can contain old commands, credentials, or malicious text.