Weights & Biases MCP server: 22 costly tools

A small tool list with the largest descriptions measured in this guide, what that costs on connection, and which two tools are responsible.

Last verified 27 August 2026 · from The wandb repository and the server's own tools/list response

This page is one of 90 in a directory of monitoring and developer tools. Each page checks what an AI coding agent can get out of the tool through its MCP server, the connector that lets the agent query the tool directly.

What is the Weights & Biases MCP server?

It is the Weights & Biases server for the Model Context Protocol. W&B tracks machine-learning experiments, and Weave tracks what large language models did, so a connected agent reaches both.

The two largest tools are query interfaces rather than fixed endpoints. query_wandb_tool executes read-only GraphQL against the W&B Models API, and query_weave_traces_tool filters and sorts Weave traces with a detail-level control.

That design explains the description sizes. A tool that accepts an open query language has to teach the model that language. The lesson lives in the description, and it is sent on every connection.

An artifact is a versioned file set such as a dataset or a model checkpoint, and a registry is where the versions a team promotes are collected.

Weights & Biases is strongest for agent observability and reaches production only where model services are involved. Its model-focused view excludes code review and CI.

Weights & Biases answers questions about models and their calls, including comparisons between training runs and evaluation summaries. It cannot describe a web service, so it complements production telemetry rather than replacing it.

What can an AI agent do with Weights & Biases?

Two calls come before the query. infer_trace_schema_tool returns the fields a trace carries, and count_weave_traces_tool returns how many match. Only then should the agent compose a query_weave_traces_tool filter.

What connecting costs before the first question

A context window is the amount of text a model can hold at once, and connecting uses some of it before the agent asks anything. Every tool the server registers sends its name, description and input schema to the model when the session starts. The tool list is charged whether or not a single tool is called.

22 tools cost between 20,920 and 34,677 tokens. Twenty-two tools cost 20,920 tokens, which is more than most servers charge for five times as many. The descriptions are the whole reason.

Descriptions average 3,271 characters, the largest measured in this guide. Schemas average 463, which is ordinary. query_wandb_tool alone carries 19,483 characters of 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 24 August 2026 from tools/list on the repository's installed package. It was run locally with a placeholder key.

Registration is static, so no W&B 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 12%. Reasoning tokens are output, spent per request. The tool list is input, charged once at connection.

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 third.

Can Weights & Biases watch your own AI agents?

Weave is the agent-observability half of this product, so this table asks what the connector reaches of it. Counted from the server's own tools/list response on 24 August 2026.

Weights & Biases MCP server limits: where answers come back incomplete

The limits below come from two tools that accept open queries and teach the model how to write them.

How to configure Weights & Biases MCP for agents

1. 1

   Budget for the descriptions, not the count

   Twenty-two tools is a small list that costs 20,920 tokens, so counting tools underestimates this connector badly.

2. 2

   Decide whether the agent may write

   create_wandb_report_tool creates reports in your workspace, and log_analysis_to_wandb writes analysis data for them.

3. 3

   Size trace queries before running them

   count_weave_traces_tool returns totals so the agent can avoid pulling a trace set that will not fit.

4. 4

   Point at the right entity

   An entity is the user or team that owns a project in W&B. list_entities_tool and query_wandb_entity_projects establish which entity and project an answer came from.

Paste this into your agent’s instructions

Two tools accept open query languages, so an answer's accuracy depends on a query the reader never sees. The block makes the agent show it.

When you answer using Weights & Biases MCP tools, state:
- The entity and project the answer came from.
- The query you ran, if you used query_wandb_tool or
  query_weave_traces_tool. Both accept open queries.
- Whether you called create_wandb_report_tool or
  log_analysis_to_wandb. Both write to the workspace.
Never state that data does not exist. State that none was
returned under the query 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?

Sometimes. W&B has a public API and a GraphQL endpoint, and query_wandb_tool is a wrapper over the latter. A script that already knows its query needs neither.

The connector earns its place on exploration across both products. Weave traces and Models runs sit behind one connection, and summarize_evaluation_tool and diagnose_run_tool do analysis the raw API does not.

For an agent whose only question is about traces, a narrower tracing connector costs far less context than 20,920 tokens.

Connect when the agent has to reason across runs and traces. Use the API directly for a known query.

Weights & Biases MCP server setup

The server installs from the repository and exposes a wandb_mcp_server entry point. One API key in WANDB_API_KEY is all it needs.

Every tool registers on connection; there is no toolset switch, so the whole 20,920-token floor applies to any session.

claude mcp add-json "wandb" '{"command":
  "wandb_mcp_server",
  "env":{"WANDB_API_KEY":"<your-key>"}}'

{"mcpServers":{"wandb":{
  "command":"wandb_mcp_server",
  "env":{"WANDB_API_KEY":"<your-key>"}}}}

[mcp_servers.wandb]
command = "wandb_mcp_server"
env = { WANDB_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.

{"mcpServers":{"wandb":{"command":"wandb_mcp_server",
  "env":{"WANDB_API_KEY":"<your-key>"}}}}

{"mcpServers":{"wandb":{"command":"wandb_mcp_server",
  "env":{"WANDB_API_KEY":"<your-key>"}}}}

{"servers":{"wandb":{"command":"wandb_mcp_server",
  "env":{"WANDB_API_KEY":"<your-key>"}}}}

{"mcpServers":{"wandb":{"command":"wandb_mcp_server",
  "env":{"WANDB_API_KEY":"<your-key>"}}}}

# Issue a W&B key without write access for the agent.
WANDB_API_KEY=<read-scoped-key> wandb_mcp_server

Read-only access, permissions and security

• Two tools write to your workspace

  create_wandb_report_tool creates reports and log_analysis_to_wandb logs data for them. Both arrive enabled.

• The report tool asks for restraint in its own description

  It states that it should be called only when the user explicitly asks for a report, which is guidance to the model rather than a control.

• Trace queries can return a great deal

  query_weave_traces_tool has a detail-level control precisely because a full trace body is large; count_weave_traces_tool exists to size a query first.

• GraphQL access is read-only by design

  query_wandb_tool describes itself as executing read-only GraphQL against the W&B Models API.

Troubleshooting

The agent uses a tenth of its window on connection

That is expected. The definitions measure 20,920 to 34,677 tokens, and the descriptions rather than the schemas carry it.

A trace query floods the context

Call count_weave_traces_tool first to size the result, then set a lower detail level on query_weave_traces_tool.

A run cannot be found

Confirm the entity and project. list_entities_tool and query_wandb_entity_projects return what the key can see, and a run in another entity looks identical to one that does not exist.

A report appeared that nobody asked for

create_wandb_report_tool is enabled by default. Its description asks the model to call it only on an explicit request, which is not a permission boundary. Use a key without write access.

Weights & Biases MCP server: Reference

What engineers report

Public discussion of the connector is thin, so the repository and the measurement carry the evidence.

“Today we only read Mcp-Session-Id from requests and never emit it in responses. As a result, our current analytics session id falls back to a hash of the API key, effectively session equals API key forever.”

Repository facts from the GitHub API, retrieved 24 August 2026. Tool counts, description lengths and schema sizes come from the server's own tools/list response on the installed package.

Should you connect an agent to Weights & Biases?

Yes. Best for a read-scoped key, with count_weave_traces_tool used before any large trace query.

Strongest fit

Teams already running W&B and Weave who want an agent to compare runs and read traces in one place.

Main advantage

Both products behind one connection, with analysis tools that go beyond what the raw API returns.

Main weakness

Twenty-two tools cost 20,920 tokens, so the tool count is a poor guide to what the connection costs.

Operational risk

Two writing tools arrive enabled, and the restraint on one of them is written into a description rather than enforced.

Frequently asked questions