To let your coding assistant read your LLM traces, add the LangWatch MCP server (https://app.langwatch.ai/mcp) to Claude Code, Cursor or Copilot. The assistant then searches the traces, opens the one that failed and reads every span, next to the code that produced it.
It also works in ChatGPT and any other client that speaks MCP. You connect it once with a URL and a browser sign-in, and then you ask “why did the last refund conversation fail?” instead of copying trace JSON into the chat.
Access stays scoped: the assistant reads only the one project you approve, with your permissions, and never sees what your redaction settings removed. To start now, go to Connect the server.
Why give the assistant the traces
In your traces the assistant finds what the code alone can’t show it: the prompt the model actually received, the tool call it made, the arguments, the error, the tokens and the time each step took.
Without the server you copy that into the chat one trace at a time, and the assistant only sees what you picked. With it, the assistant searches for the failures itself, reads them in a compact digest made for a model, and checks the next run after it changes the code.
1. Connect the server
The server is hosted at https://app.langwatch.ai/mcp. In Claude Code:
Then run /mcp, pick langwatch and choose Authenticate. A browser window opens: sign in to LangWatch and approve the project the assistant may read.
In Cursor, Copilot, ChatGPT and other clients you add a remote MCP server with the same URL and sign in the same way. LangWatch MCP setup has the steps for each client, the personal API key option for clients that can’t sign in, and a local version that runs with npx.
Your app needs to send traces to LangWatch first. If it doesn’t yet, ask the assistant to “instrument this project with LangWatch”: the server gives it the integration docs for your framework.
2. Ask for the traces that went wrong
Ask the way you would ask a colleague, and the assistant picks the tools:
search_traces matches the text of inputs and outputs, trace names and span names, within a date range and filters such as labels or user id. get_trace opens one trace by id, and for traces from the last 90 days a unique prefix of 8 or more characters is enough, like a git commit. Both return a digest: the span tree with timings, inputs, outputs and errors, short enough to fit in the assistant’s context.
3. Fix it and check the next run
Because the assistant has your code open, it can go from the trace to the line that caused it. A typical loop:
- The assistant reads the failing trace and finds the step that went wrong, for example a tool call with a missing argument.
- It edits the prompt or the code and runs the agent again.
- It searches for the new trace and compares it with the old one, span by span.
You see the reasoning in the chat and the change in your diff, so you decide what stays.
4. Count across many traces
For questions about many traces at once, the run_query tool runs a read-only SQL query (LangWatchQL) over the project’s traces, spans and evaluations:
The assistant reads the available views first with discover_schema, then writes the query, and the answer comes back as a table.
5. Turn a bad trace into a test
When you have found a failure, keep it from coming back in the same chat:
- Add it to a dataset. The assistant can add the trace’s input and the expected output as a row in a dataset, which experiments and evaluations then run against.
- Write a scenario test. The assistant can write a simulated-user test that starts the way that user did, so the same mistake fails a test in CI.
- Mark the trace. It can add a thumbs down and a comment to the trace for whoever reviews it next.
What the assistant can see
With the sign-in, the assistant reads only the project you approve, with your permissions, and only what your redaction settings kept. What it does read, inputs and outputs included, goes to the assistant’s model provider as part of the chat. The limits in detail:
- Only the approved project. The sign-in approves one project, and the assistant reads only that project’s traces. A personal API key instead reaches every project its scope covers, which can be more than one.
- Your redaction applies. The server returns traces as LangWatch stored them, so anything your data privacy settings redacted or dropped is not there to read.
- Your access applies. The assistant acts as you, so it can’t read or change anything your LangWatch role can’t.
Questions about the LangWatch MCP server
Which assistants work with the LangWatch MCP server
Claude Code, Claude Desktop, Cursor, GitHub Copilot in VS Code, ChatGPT and any client that supports remote MCP servers over Streamable HTTP. The local version runs in any client that starts MCP servers as a command.
Does the LangWatch MCP server work with self-hosted LangWatch
Yes. Use your own host followed by /mcp as the URL.
Can the LangWatch MCP server change things, or only read
It can also create and update prompts, evaluators, datasets, scenarios and online evaluations. The full list is in the tools reference. Whether you approve each call depends on your client’s permission settings, so check them before you let it write.
Do my traces need a particular framework to use the LangWatch MCP server
No. Anything that sends traces to LangWatch works, through the Python or TypeScript SDK or plain OpenTelemetry.
Next, connect the server with the command in step 1 and ask your assistant for the last failed trace, or read LangWatch MCP for every tool and setting. Last modified on November 10, 2026
