fetch_runs and get_thread_history page by character budget rather than row count, so a long thread comes back in readable pages instead of one response that swamps the context. Datasets, examples and experiment projects sit beside the runs, which is exactly what you need when the open question is whether a regression is in the chain or in the eval set that judges it.
A read path into a LangSmith workspace: runs and thread history from tracing projects, the prompt registry, datasets and their examples, experiment projects, and organisation billing usage. Four of the fifteen tools return written instructions rather than data — they cover the write paths.
- Runs fetched from one or more projects, filtered by run type, root-only, error state, trace ID, and trace or tree filters
- One page of a conversation thread's message history, with long strings cut to a preview length you set
- Character-budget pagination on both, so a long trace arrives as pages you can iterate rather than one response that will not fit
- Projects listed at simplified or full detail, filtered by name or by the dataset they reference
- The prompt registry: prompts listed with a public or private filter, or one fetched by its exact name
- Datasets fetched by ID, name or metadata, and their examples with filters, splits and attachments
- Experiment projects for a named dataset, which is where evaluation runs are compared
- Organisation billing usage as trace counts, with workspace names filled in beside each workspace UUID
A LangSmith workspace whose projects, prompts and datasets you can read, and uv on your PATH.
One command plus a key — uvx langsmith-mcp-server, then supply credentials