LangSmith/Q/What can the LangSmith Remote MCP do that the LangSmith CLI can't, and vice versa?

• What can the LangSmith/MCP/Remote do that the LangSmith/CLI can't, and vice versa?

  • AI Answer

    • Answer/Official from LangSmith Remote MCP, LangSmith CLI, LangSmith MCP Server, Sandbox CLI, and the langsmith-cli README.
    • Short answer: yes, in both directions, and the CLI is the larger surface. Both read projects, traces, runs, threads, datasets, examples, and experiments — LangChain's own docs say the CLI "supports the same projects, traces, runs, datasets, experiments, and threads as the MCP server." Beyond that overlap, the Remote MCP uniquely offers typed prompt and billing tools plus zero-install OAuth from inside a client, while the CLI uniquely offers writes, exports, sandboxes, evaluators, Hub repos, agent tracing setup, and a raw REST escape hatch.

    • Only the Remote MCP

      • Typed prompt tools. list_prompts and get_prompt_by_name read LangSmith/PromptHub prompts and templates. The CLI has no prompt command group at all — it can only reach prompts through langsmith api.
      • Billing usage. get_billing_usage returns organization trace counts for a date range, optionally filtered by workspace. There is no CLI billing command.
      • Nothing to install. A client connects with a URL and an OAuth login; no binary, no PATH, no self-update. That matters for hosted or sandboxed agents that cannot install software.
      • Model-driven tool selection. Tools arrive in the model's context with schemas, so the agent picks and fills them itself instead of an agent-authored shell command being parsed and executed.
      • Character-budget pagination. get_thread_history and fetch_runs page by character count (max_chars_per_page, preview_chars) so a response fits a context window. The CLI's equivalent is --limit / --offset plus detail flags — item-based, not size-based.

    • Only the CLI

      • Sandboxes. langsmith sandbox builds snapshots from Docker images, boots VMs, runs exec, opens interactive consoles, and tunnels TCP ports into a sandbox. No MCP tool touches sandboxes.
      • Evaluator management. langsmith evaluator upload uploads code evaluators as offline (dataset) or online (project) rules, with --sampling-rate and --replace; evaluator list and evaluator delete round it out. Compare LangSmith/Evaluator and LangSmith/Evaluator/Online.
      • Real writes and deletes. dataset create / delete / upload, example create / delete. The MCP's create_dataset, update_examples, push_prompt, and run_experiment are documentation-only tools — they return instructions on how to do the thing, not the thing itself.
      • Bulk export to disk. trace export, run export (JSONL), dataset export, and the global -o <path>. MCP tools return payloads into a conversation, capped by the character budget.
      • LangSmith/Engine issues. langsmith project issues list --project <name> lists Engine-detected issues.
      • Raw REST escape hatch. langsmith api is a gh api-style authenticated wrapper over the whole LangSmith REST API, with langsmith api ls / api info to browse the OpenAPI spec. This makes the CLI a strict superset in reach: anything the API exposes, including endpoints behind MCP tools, is callable.
      • Hub repos. langsmith hub init / push / pull / list / get / delete manage versioned agent and skill repos on the LangSmith Hub.
      • Tracing coding agents. langsmith trace setup claude and langsmith trace setup codex write Claude/Code and Codex config and install the tracing plugin so the agent's own sessions land in a LangSmith project.
      • Shell-native ergonomics. --format pretty tables, the filter DSL via --filter, --show-hierarchy run trees, --min-latency / --min-tokens filters, named profiles, and pipes into jq.
      • Works where MCP OAuth doesn't. The docs explicitly route Codex users to the CLI because Codex's OAuth flow is incompatible with the Remote MCP.

    • Where they differ on deployment and auth

      • Both hit the same OAuth authorization server, so langsmith auth login and an MCP client's /mcp login are the same identity flow.
      • Self-hosted: the Remote MCP requires LangSmith v0.16+ and an Ed25519 signing JWKS before /api/mcp stops returning 404. The CLI reaches any self-hosted instance with LANGSMITH_ENDPOINT and an API key, though langsmith auth login is Cloud-only.

    • Practical read

      • Use the Remote MCP when an agent should discover LangSmith capabilities on its own with no install step, or when prompts and billing are the target. Use the CLI when the task involves writing, exporting, sandboxes, evaluators, or anything scripted — and note that langsmith api means the CLI can eventually do everything the MCP can, while the reverse is not true.