W&B MCP Server
An agent reads Weights & Biases runs, metrics and Weave traces and builds reports from plain-language questions
- 72 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The W&B key opens every project and entity visible to the key's owner
- `create_wandb_report_tool` and `log_analysis_to_wandb` create reports and runs in the account
- Weave traces can hold user prompts and model answers, and they enter the agent's context
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/w-b-mcp-serverDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The W&B key opens every project and entity visible to the key's owner
- `create_wandb_report_tool` and `log_analysis_to_wandb` create reports and runs in the account
- Weave traces can hold user prompts and model answers, and they enter the agent's context
This is third-party code. Review the repository files before installing.
What it does
The official Weights & Biases MCP server gives an agent the data in W&B Models and Weave. The agent finds runs, compares configurations and training curves, and reads LLM call traces and artifact versions in model registries. There are more than twenty tools. `query_wandb_tool` reads projects and runs, `get_run_history_tool` returns bounded time series for several metrics, `compare_runs_tool` sets runs side by side, and `diagnose_run_tool` explains why a run diverged. For Weave traces the agent first calls `infer_trace_schema_tool` to see the fields, then `query_weave_traces_tool` takes the columns it needs at a chosen level of detail. `create_wandb_report_tool` builds a report with charts, and `search_wandb_docs_tool` searches the W&B documentation. The server connects at https://mcp.withwandb.com/mcp or runs locally. W&B Dedicated and Self-Managed installations use the address of their own instance. The `WANDB_MCP_ACCESS_MODE=read-only` variable removes the write tools.
Who it is for. ML engineers and researchers who track experiments in W&B, and teams debugging LLM applications through Weave traces.
Good fit when
- Among a hundred runs you need the five with the best `eval/accuracy`, with their configurations compared
- Training diverged: the agent sets the loss curves of two runs side by side and names the step where they split
- An LLM application keeps failing, and the last hundred Weave traces should show the most common errors
- After a series of experiments, a W&B report with loss charts is built from the chat
- Model versions in the registry differ, and the agent shows what changed between two artifact versions
Not a fit when
- Experiments live in another tracker: the server reads W&B only
- Metrics have to be logged from training code: that is the job of the `wandb` library, while the server writes only reports and summary metrics of an analysis
- The answer must cover the whole project: the tools return bounded samples, and the `returned_count` and `has_more` fields show what was left out
- You do not want to hand the agent a key with access to every project: the server acts with the key's rights
Example request
Show the five best runs in my-team/my-project by `eval/accuracy` and compare their configurations.MCP
- Transport
- http, stdio
- Authentication
- API key
Environment variables
WANDB_API_KEYrequired, secret- A W&B API key from wandb.ai/authorize.
WANDB_BASE_URL- Address of a W&B Dedicated or Self-Managed instance for a local run.
WANDB_MCP_ACCESS_MODE- The value `read-only` removes the write tools.
Limitations
The server does not work anonymously: it needs a W&B key from wandb.ai/authorize. The tools return bounded samples, so a broad question such as “which run is best” gets an incomplete answer, and the question is better narrowed to one metric and project.
The write tools create objects in the account: a report or a run with analysis results. Some tools (ARIA, Agent Lens, Weave agents) are switched on by separate profiles and are unavailable in a Dedicated deployment.
Whether the hosted server is reachable from Russia was not checked for this card.
How to disable. Remove the server with `claude mcp remove wandb`. To keep it but forbid writing, run it locally with `WANDB_MCP_ACCESS_MODE=read-only`.
FAQ
How do I connect W&B MCP Server to Claude Code?
With one command: `claude mcp add --transport http wandb https://mcp.withwandb.com/mcp --scope user --header "Authorization: Bearer <key>"`. Get the key at wandb.ai/authorize. Name the entity and project in your requests, otherwise the agent has to guess.
Can the agent be kept from writing to W&B?
Yes. On a local server, `WANDB_MCP_ACCESS_MODE=read-only` removes every write tool, including `create_wandb_report_tool` and `log_analysis_to_wandb`. `query_wandb_tool` reads data in any mode.
Does the server work with W&B Dedicated and Self-Managed?
Yes. If the operator enabled MCP in the chart, the server address looks like `https://<your-instance>/mcp`. A local server is pointed at the instance with `WANDB_BASE_URL`.
Why is the answer incomplete when I ask for the best run?
The tools return bounded samples. The `total_count`, `returned_count`, `has_more` and `truncated` fields show what stayed outside the answer. Narrow the question: name the metric and project, for example “which run had the highest f1”.
Which agents does the server work with?
The documentation has separate instructions for Cursor, Claude Code, Claude Desktop, Codex, Gemini CLI, VS Code, the OpenAI Responses API and Le Chat.
Discussion
No questions or reviews yet.
Related
MCP serversMCP server for SAP BW/4HANA and BW 7.5 on HANA: an agent reads and changes modeling objects, runs queries and watches data loads.
A lightweight database workspace with an MCP server: the agent reads the schema, runs queries and explores data in PostgreSQL, MySQL and more
An MCP server with a toolbox for global briefs: markets, SEC filings, conflict, climate and news from open sources
MySQL MCP Server Pro
mysql_mcp_server_pro
A MySQL MCP server that runs SQL under access roles, finds tables by comments and analyzes locks, indexes and query plans
An agent queries GreptimeDB metrics, logs and traces with SQL, TQL and RANGE, with write protection and data masking
A local Microsoft MCP server for Power BI semantic models: an agent creates tables, DAX measures and relationships in Desktop, Fabric and PBIP files
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| W&B MCP Server (this card) | yes | API key | 72 | |
| Langfuse Skills | yes | 339 | works with traces in Langfuse and does not read W&B data | |
| LangSmith skills | no | 166 | works with LangSmith and does not see W&B experiments | |
| Arize Phoenix | yes | OAuth | 12,268 | LLM observability in Arize Phoenix, a different service and different data |
README in short
Installation in the documentation splits into the hosted server, which is recommended, and a local run. Each client has its own block with a command or a configuration snippet. Version 0.4 added structured experiment queries, tool profiles, a read-only mode and bounded response sizes.