transcripts-mcp

Search your Cursor, Claude Code, and Codex conversations from any MCP client.

transcripts-mcp reads session transcripts already on your machine. Use it to find a previous implementation, recover the reasoning behind a decision, or bring context from one coding tool into another.

Website · Setup · Tool reference · Contributing

What it does

• Browse sessions across all three tools, filtered by project or date.

• Read conversations as normalized messages, with a limit on how much is returned.

• Search immediately with fuzzy, plain-text, or regex matching. No index required.

• Rank results with a local full-text index and optional semantic search.

For example, ask your connected client:

Find the Claude Code conversation where we added authentication to this project.

Read my most recent Cursor session and summarize what remains to be done.

Search my Codex transcripts for the database migration error.

The server returns transcript text for the current client to read. It does not restore a previous session's model state.

Related MCP server: transcript-search

Install

Requires Node.js 24 or later for the npm launcher. Add the command below to your MCP client using one of the configurations in the next section:

npx -y transcripts-mcp

You can also launch with pnpm dlx transcripts-mcp or bunx --bun transcripts-mcp.

The npm and pnpm launchers use a platform binary with grep and full-text search. If that binary is unavailable, they look for Bun on PATH. Optional semantic search requires running under Bun; see Search.

The server communicates over stdio and is intended to be launched by an MCP client.

Configure

Choose your client below, save the configuration, and restart the client.

Cursor

Add to ~/.cursor/mcp.json for all projects, or .cursor/mcp.json for one project:

{
  "mcpServers": {
    "transcripts": {
      "command": "npx",
      "args": ["-y", "transcripts-mcp"]
    }
  }
}

Claude Code

Register the server for your user account:

claude mcp add --scope user transcripts -- npx -y transcripts-mcp

For project scope, add the same mcpServers configuration shown above to .mcp.json at the repository root.

Codex

Add to ~/.codex/config.toml:

[mcp_servers.transcripts]
command = "npx"
args = ["-y", "transcripts-mcp"]

Cursor plugin

The repository also includes a Cursor plugin in distribution/plugin. For local development on macOS or Linux, symlink it into Cursor's local plugin directory:

ln -s /path/to/transcripts-mcp/distribution/plugin ~/.cursor/plugins/local/transcripts-mcp

The plugin includes an MCP configuration pinned to a package version.

Claude Code plugin

Install the MCP server and transcript-search skill together from the repository's marketplace. Run these commands inside Claude Code:

/plugin marketplace add Stormix/transcripts-mcp
/plugin install transcripts-mcp@stormix-plugins

The Claude Code and Cursor plugins share the same version-pinned MCP configuration and skill in distribution/plugin. Requires Node.js 24 or later for the npm launcher.

For local development, run from the repository root:

claude --plugin-dir ./distribution/plugin

Check /mcp for the transcripts server, then try /transcripts-mcp:transcript-search. Run /reload-plugins after changing the plugin configuration.

See the Claude Code plugin docs for development and troubleshooting.

Tools

Bold inputs are required. Provider IDs are cursor, claude-code, and codex. Date filters use ISO-8601 timestamps.

Result limits default to 50 for session listing and grep, 200 messages for transcript reads, and 20 for indexed search. Maximums are 200, 1,000, and 100 respectively.

Search

Search without an index

grep_transcripts searches raw JSONL files and returns adapter-normalized results. It supports fuzzy (the default), plain, and regex modes.

Search uses @ff-labs/fff-bun. If the native library cannot load, the server falls back to a streaming scan.

Full-text search

Call build_index, then use search_transcripts with mode: "fts" (the default). Results are ranked with SQLite FTS5's BM25 scoring over normalized message text.

Run build_index again to pick up transcript changes. Set full: true to rebuild from scratch.

Semantic search

Run the server with Bun:

{
  "mcpServers": {
    "transcripts": {
      "command": "bunx",
      "args": ["--bun", "transcripts-mcp"]
    }
  }
}

Call build_index with semantic: true, then search with mode: "hybrid". Hybrid search combines full-text and vector results using reciprocal rank fusion.

The first semantic build downloads the all-MiniLM-L6-v2 model. Embeddings are computed locally. The compiled platform binary does not include the semantic engine.

Configuration

Transcript directories are discovered automatically. Set these environment variables in your MCP client's server configuration to override the defaults:

Privacy and local storage

The server reads transcripts without modifying them. Search and embedding computation run locally; the server does not upload transcript content.

Results are returned to the MCP client that requests them. That client may send retrieved text to its model provider according to its own settings.

Indexed search stores message text and, when enabled, embeddings in the local SQLite index. Semantic search also downloads and caches model files on first use.

From source

Requires Bun >= 1.2, Node.js >= 24, and pnpm 11.

git clone https://github.com/Stormix/transcripts-mcp.git
cd transcripts-mcp
pnpm install
pnpm start

To connect an MCP client to your checkout, use an absolute path to the server entry point:

{
  "mcpServers": {
    "transcripts": {
      "command": "bun",
      "args": ["/absolute/path/to/transcripts-mcp/apps/mcp/src/index.ts"]
    }
  }
}

On Windows, use forward slashes or escaped backslashes in JSON paths. The server writes JSON-RPC to stdout and logs to stderr.

Contributing

See CONTRIBUTING.md for development setup, checks, and instructions for adding a transcript adapter.

The TypeScript monorepo is organized around three libraries: packages/core defines the transcript model, packages/adapters reads each provider's format, and packages/search implements search. The stdio server lives in apps/mcp; the published launcher lives in packages/cli.

AI-assisted contributions follow the AI Policy. Please use SECURITY.md to report vulnerabilities and follow the Code of Conduct.

License

MIT