LangSmith MCP server: read traces, prompts and datasets

LangSmith now hosts its MCP server, so the archived package is no longer the route. What the 15 tools reach, which client cannot connect, and what the older package is still for.

Checked 3 October 2026 · from LangSmith's documentation, the live endpoint's OAuth metadata, PyPI, GitHub and the package's own tools/list response

One of 91 guides to monitoring and developer tools. Each covers what an AI coding agent can get out of the tool through MCP, the protocol that lets the agent query the tool directly.

What is the LangSmith MCP server?

It is LangChain's server for the Model Context Protocol. LangSmith captures the runs of your own LLM applications, and these tools let an agent read those runs back.

The surface splits into five areas: prompts, traces and runs, datasets and examples, experiments, and billing usage. That last one is unusual among agent-observability connectors, which mostly stop at telemetry.

The Remote MCP runs inside LangSmith, so there is nothing to install for Cloud users. The standalone package remains for self-hosted LangSmith older than v0.16 and for teams that want to run the server themselves.

LangSmith covers agent observability broadly, reaches part of the way into release through prompt management and into cost through billing usage, and reaches nothing in code review or production observability.

For an agent, LangSmith answers questions about how your LLM application behaved and what you fed it, and nothing about the machines underneath. An agent connected here can pull the runs behind a bad answer and the dataset a prompt was tested against. It cannot tell you the database was slow, so this connector pairs with an infrastructure one.

What can an AI agent do with LangSmith?

What the tool definitions cost in context

A client that loads every tool definition at the start of a session spends this much context before the first question. Claude Code defers definitions by default through tool search, and loads a tool's definition when the agent needs it. The method below gives the version and date measured.

A context window is the amount of text a model can hold at once. A client that loads every tool definition upfront sends each tool's name, description and input schema to the model when the session starts. That list is charged whether or not a single tool is called.

Measured on the standalone package, LangSmith is verbose for its size: 15 tools cost about 5,458 tokens. Arize Phoenix registers 27 tools and costs less. LangSmith's tool descriptions average 1,120 characters. 1,120 characters is long for a tool description.

Encodings: GPT-5.6, GPT-5.5 o200k_base; GPT-4, GPT-3.5 cl100k_base; Claude Haiku 4.5 to Opus 5 count_tokens. Counted 21 August 2026 from tools/list on the standalone package langsmith-mcp-server 0.1.1, run locally with a placeholder key. The Remote MCP shares its tool implementations, but its own tools/list was not counted, so its definitions may serialize differently.

Registration is static, so no LangSmith account was involved. Figures cover name, description and input schema.

The two models in column one share the o200k_base encoding, so they get one count. The spread is formatting, about 22%. Reasoning tokens are output, spent per request. The tool list is billed as input whenever the client loads it.

The Claude figures are measured rather than encoded. Anthropic publishes no local tokenizer. So the definitions were sent to the API's count_tokens endpoint, and a no-tools baseline was subtracted. The range covers two tokenizer generations rather than two serializations. Haiku 4.5 and Opus 4.6 sit at the low end. Opus 5 and Sonnet 5 sit at the high end. The gap between them is about a quarter.

Can LangSmith watch your own AI agents?

LangSmith is built for tracing agents, so the table below asks which of those signals the 15 tools expose.

LangSmith MCP server limits: where answers come back incomplete

The first limit below is about which server you connect; the other two are about what the tools return.

How to configure LangSmith MCP for agents

1. 1

   Use the Remote MCP

   The standalone package is deprecated and will not change, while the hosted server follows LangSmith releases.

2. 2

   No allowlist is needed for safety

   Every tool reads. The four documentation-only tools are worth disabling for context rather than for risk: they return long markdown pages an agent rarely needs.

3. 3

   Pick the right region

   EU, APAC and AWS US accounts use eu., apac. and aws. prefixes on api.smith.langchain.com/mcp.

4. 4

   Mind the workspace

   OAuth sessions follow your user's workspace permissions. With an API key, tools that accept workspace_id can target another workspace; otherwise the key's own workspace is used.

Paste this into your agent’s instructions

LangSmith's quiet failure is scope: a run search in the wrong project, or a thread with no identifier, both return nothing. The block makes the agent name what it searched.

When you answer using LangSmith MCP tools, state:
- The project and workspace you searched.
- The time range, and how many runs you actually retrieved.
- Whether a thread identifier existed. get_thread_history
  returns nothing when the application never set one, which
  is not the same as a conversation with no turns.
Never state that data does not exist. State that none was
returned under the constraints above.
Write your answer in ASD-STE100 Simplified Technical
English. Use short sentences with one idea in each.

Do you need the MCP server at all?

The LangSmith CLI signs in through the same OAuth server and covers the same projects, traces, runs, datasets, experiments and threads. LangSmith recommends it where the Remote MCP cannot connect, such as Codex.

The LangSmith REST API and SDK cover runs, prompts and datasets, and they are the route for the writes the MCP tools only describe.

The connector's advantage is tool descriptions detailed enough that an agent picks the right call unprompted, which is the same verbosity that makes the package's definitions cost 5,458 tokens.

Connect the Remote MCP for exploring recent runs and prompts. Use the SDK or CLI for anything that changes data.

LangSmith MCP server setup

LangSmith Cloud users connect to https://api.smith.langchain.com/mcp, or the eu., apac. or aws. regional host. OAuth 2.1 with dynamic client registration means no key or client ID to provision.

For scripts, send a LangSmith API key as X-Api-Key. The deprecated standalone package uses a different header, LANGSMITH-API-KEY, when run over HTTP.

claude mcp add --transport http -s user langsmith https://api.smith.langchain.com/mcp

{"mcpServers":{"LangSmith":{"url":"https://api.smith.langchain.com/mcp"}}}

[mcp_servers.langsmith]
command = "uvx"
args = ["langsmith-mcp-server"]
env = { LANGSMITH_API_KEY = "<your-key>" }

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

{"servers":{"langsmith":{"type":"http","url":"https://api.smith.langchain.com/mcp"}}}

const client = await createMCPClient({
  transport: {
    type: "http",
    url: "https://api.smith.langchain.com/mcp",
    headers: { "X-Api-Key": process.env.LANGSMITH_API_KEY },
  },
});

LANGSMITH_API_KEY=<your-key> \
LANGSMITH_ENDPOINT=https://langsmith.example.com \
uvx langsmith-mcp-server

Read-only access, permissions and security

• No tool writes

  Four of the fifteen are named like writes, but push_prompt, create_dataset, update_examples and run_experiment are labeled Documentation-only in the README and implemented as zero-argument functions that return a markdown page of SDK examples. They take no parameters and reach nothing.

• The naming is the hazard, not the behavior

  An operator reading the tool list reasonably assumes push_prompt pushes a prompt. It returns instructions for pushing one. Budget for the confusion, not for the write.

• Nothing here spends money

  run_experiment returns documentation, so an agent calling it repeatedly costs context and nothing else.

• Access follows the identity

  An OAuth session sees what your LangSmith user may see. An API key acts as the user who owns it, within that key's workspace and permissions.

Troubleshooting

No runs found in a project you know has traffic

Check the project name and the workspace. LANGSMITH_WORKSPACE_ID scopes results, and a query against the wrong workspace returns an empty result with no explanation.

A conversation has no thread history

get_thread_history needs a thread identifier that your application sets. Without one there is nothing to group, and the response looks the same as an empty conversation.

Codex CLI logs in but every call fails

LangSmith documents this as a known incompatibility: Codex omits the OAuth resource parameter, so the token is not bound to the server. Use the LangSmith CLI from Codex, or the standalone package with an API key.

401 with an API key

The Remote MCP reads X-Api-Key. The standalone package's HTTP mode reads LANGSMITH-API-KEY instead, so a header copied between them fails.

The self-hosted /api/mcp endpoint returns 404

The Remote MCP stays inert until a signing JWKS is configured. Generate an Ed25519 key set and pass it as config.signingJwks; RSA keys are rejected.

LangSmith MCP server: Reference

What engineers report

Signals for the Remote MCP are LangSmith's own documentation. The standalone package left a public record: 131 stars, 11 PyPI releases and an archive date of 14 August 2026.

“exceeds maximum allowed tokens (25000). Please use pagination, filtering, or limit parameters to reduce the response size.”

Remote MCP facts from LangSmith's documentation and the endpoint's OAuth metadata, retrieved 3 October 2026. Package facts from the GitHub API and PyPI on the same date. Tool counts come from the package's own tools/list response, which matches the documented Remote MCP list.

Should you connect an agent to LangSmith?

Yes. Best for LangSmith teams debugging traces and checking experiments from Claude Code or Cursor.

Strongest fit

Yes, if you already trace with LangSmith and want an agent reading runs, prompts and experiments; connect the hosted Remote MCP.

Main advantage

Hosted with OAuth across four Cloud regions, and it reaches billing usage, which most agent-observability connectors omit.

Main weakness

Four of the 15 tools only return documentation, and Codex CLI cannot connect to the hosted server.

Operational risk

There are no writes to gate. The four tools that read as writes return documentation, so the connector is read-only in practice even though nothing advertises it as such.

Frequently asked questions