LangSmith Tracing for Claude Code

Sends your Claude Code conversations to LangSmith so you can read back what the agent actually did.

What you need

• On a Mac, nothing. The plugin carries its own build and runs it directly.
• Everywhere else, including Windows, Node.js 20 or newer.
• A LangSmith account and API key.

If the carried Mac build cannot start, the plugin falls back to Node rather than dropping the trace, so Node is worth having on a Mac too.

Install

Run these inside Claude Code:

/plugin marketplace add langchain-ai/langsmith-claude-code-plugins
/plugin install langsmith-tracing@langsmith-claude-code-plugins
/reload-plugins

Updates are off by default for this marketplace, so turn on auto-update in /plugin under Marketplaces, or run claude plugin update langsmith-tracing@langsmith-claude-code-plugins. Either way the new version loads in your next session.

Turn on tracing

Tracing is off until you give it a key and switch it on. Write both to ~/.claude/langsmith.json:

{ "enabled": true, "api_key": "lsv2_pt_...", "project": "claude-code" }

Get a key from smith.langchain.com under Settings then API Keys. Send a message, then look for it in the claude-code project.

A project can carry its own settings in /.claude/langsmith.json, and a setting in your shell beats any file. Advanced configuration lists the other locations.

Check a repository's tracing settings before you trust it. A repository you clone can switch tracing on and send your conversation to someone else's server. Review these files in an unfamiliar repository. Check .claude/settings.json and .claude/settings.local.json too.

What gets traced

Each model call carries the conversation so far and the assistant's reply. You also get the model name, provider and token counts. Tool calls come with their inputs and outputs. Skill calls record which skill ran so you can count usage per skill. A subagent, meaning a helper Claude Code spawns to work on its own, appears nested under the turn that started it.

A tool call that touches a path is labelled with the repository that path is in, rather than the one you started in, so a session spanning several repositories labels each call correctly. Those calls carry that repository's name, branch and commit, plus the name git would put on a commit there, which is usually your global git name, so treat it as personal data. A path in no repository carries none of them.

Cancel a turn and it still uploads, marked interrupted. A subagent uploads only once it finishes, so cancelling one leaves its own steps out.

Run /langsmith-tracing:trace to get a link to the current conversation's trace. It does not start a model turn or change anything.

Hide one conversation

To keep one conversation's content out of LangSmith, run these inside Claude Code:

/langsmith-tracing:mute
/langsmith-tracing:unmute

Muting keeps the shape of the conversation and leaves out what was said. It starts from your next turn, so wait for the confirmation before sending anything sensitive.

Muting changes only what reaches LangSmith. Claude Code still reads and remembers everything locally. Earlier uploads are not deleted.

To mute every conversation, set defaultMuted to true. A conversation you muted or unmuted by hand keeps that choice.

Change a setting