Weights & Biases MCP integration for AI agents.

Connect AI agents to Weights & Biases through 22 structured actions, including count weave traces, create report, and compare artifact versions. Review authe…

ACTION CONTRACTS
22
AUTHENTICATION
API key
SOURCE STATUS
Provider-backed

CATALOG SOURCE REVIEWED AUGUST 23, 2026 / ACTION NAMES AND SCHEMAS DERIVED FROM WORKING MACHINES PROVIDER SOURCE

CAPABILITY PROFILE

What the Weights & Biases integration exposes

The Weights & Biases Agent App exposes 22 provider-backed actions for ai, analytics, and developer tools work. Its current contract lets an authorized agent count matching classic weave traces and root traces without returning trace payloads, create a w&b report with narrative text, plots, and configurable panels, and compare metadata, aliases, tags, lineage, sizes, and optionally files for two w&b artifact versions. These operations are called through Working Machines as typed capabilities rather than through browser navigation or copied UI steps.

A connection uses API key. Before execution, the agent can inspect the selected action, its required fields, declared scopes, and expected output contract. Provider credentials remain inside the Working Machines runtime; the calling agent receives the capability and its structured result, not the underlying secret.

For reliable operation, start with the narrowest action that satisfies the task, resolve stable provider identifiers before changing state, and validate the returned object or status after execution. Availability still depends on the connected Weights & Biases account, granted provider permissions, workspace policy, region, plan, and upstream API behavior.

VERIFIED ACTION SAMPLE

Real Weights & Biases capabilities.

Showing 12 of 22 actions. Risk labels are conservative signals based on operation names, not substitutes for provider documentation or runtime policy.

count_weave_tracesREAD

Count matching classic Weave traces and root traces without returning trace payloads.

INPUTS: entity_name / project_name / filters

create_reportWRITE

Create a W&B report with narrative text, plots, and configurable panels.

INPUTS: entity_name / project_name / title / description / markdown_report_text

compare_artifact_versionsREVIEW

Compare metadata, aliases, tags, lineage, sizes, and optionally files for two W&B artifact versions.

INPUTS: artifact_name_a / artifact_name_b / type_name / include_file_diff / max_file_diff_entries

compare_runsREVIEW

Compare configuration, summary metrics, metadata, and optionally metric history for two W&B runs.

INPUTS: entity_name / project_name / run_id_a / run_id_b / include_history_overlap

diagnose_runWRITE

Inspect loss metrics and run state to diagnose the training health of a W&B run.

INPUTS: entity_name / project_name / run_id / loss_key / val_loss_key

get_artifact_detailsREAD

Get metadata, lineage, and optionally files for a W&B artifact version.

INPUTS: artifact_name / type_name / include_files / max_files

get_run_historyREAD

Retrieve sampled time-series metric history for a W&B run.

INPUTS: entity_name / project_name / run_id / keys / samples

infer_trace_schemaREVIEW

Sample classic Weave traces to discover field paths, types, and frequent values.

INPUTS: entity_name / project_name / sample_size / top_n_values

list_artifact_versionsREAD

List versions of a W&B project artifact or registry collection.

INPUTS: collection_name / entity_name / project_name / registry_name / organization

list_automationsREAD

List W&B Automations that trigger on artifact, run-state, or run-metric events.

INPUTS: entity / name / max_items

list_entitiesREAD

List W&B user and team entities accessible with the configured API key.

list_integrationsREAD

List Slack and webhook integrations available as W&B Automation targets.

INPUTS: entity / kind / max_items

INPUT CONTRACTS

Know what the action needs before it runs.

count_weave_traces

entity_nameREQUIRED
The W&B entity or team name.
project_nameREQUIRED
The W&B project name.
filtersOPTIONAL
Optional provider-native filters.

create_report

entity_nameREQUIRED
The W&B entity or team name.
project_nameREQUIRED
The W&B project name.
titleREQUIRED
The report title.
descriptionOPTIONAL
Optional report description.
markdown_report_textOPTIONAL
Markdown narrative for the report.

compare_artifact_versions

artifact_name_aREQUIRED
The first fully qualified artifact version name.
artifact_name_bREQUIRED
The second fully qualified artifact version name.
type_nameOPTIONAL
Optional artifact type name.
include_file_diffOPTIONAL
Whether to include file-level differences.
max_file_diff_entriesOPTIONAL
Maximum file-difference entries to return.

compare_runs

entity_nameREQUIRED
The W&B entity or team name.
project_nameREQUIRED
The W&B project name.
run_id_aREQUIRED
The first W&B run ID.
run_id_bREQUIRED
The second W&B run ID.
include_history_overlapOPTIONAL
Whether to compare overlapping sampled metric history.

PROVIDER-SPECIFIC WORKFLOWS

Jobs this Agent App can support

Inspect Count Weave Traces

Count matching classic Weave traces and root traces without returning trace payloads. Use this as a bounded discovery step, retain the returned identifier, and avoid expanding the read beyond the task's stated scope.

count_weave_traces

Control Create Report

Create a W&B report with narrative text, plots, and configurable panels. Resolve the target first, present material changes for confirmation, and make retries idempotent where the provider supports it.

create_report

Verify with Get Artifact Details

Get metadata, lineage, and optionally files for a W&B artifact version. Compare the returned provider state with the intended outcome and preserve stable IDs or canonical links in the run record.

get_artifact_details

SAFETY BOUNDARY

Operate Weights & Biases with explicit limits

  • Authorize Weights & Biases with API key and grant only the provider access required by the selected actions.
  • 2 actions are change-capable by name. Confirm the target identity and material parameters before allowing a write.
  • Do not infer permission from catalog visibility. Workspace policy, connection identity, and upstream authorization still govern execution.
  • No provider scope string is declared on the sampled actions. Verify the connected account's actual permissions in Weights & Biases rather than assuming unrestricted access.
  • After a call, inspect the structured result and execution record before reporting that the Weights & Biases task completed successfully.

CONNECTION MODEL

API key

A W&B API key sent to the MCP server as an Authorization Bearer token. Create or copy a key at https://wandb.ai/authorize.

Working Machines stores provider credentials behind the execution boundary. An agent can use an authorized connection identity, but catalog discovery alone does not reveal OAuth tokens, API keys, or provider secrets.

READ SIGNALS
13
WRITE SIGNALS
2
HIGH IMPACT
0
REVIEW SIGNALS
7

EVIDENCE AND AVAILABILITY

Provider reference

Action names, input fields, authentication types, and counts on this page are generated from the Working Machines provider catalog. Provider behavior, quotas, object semantics, account eligibility, and regional availability remain governed by Weights & Biases.

Official Weights & Biases website

ONE CONNECTION. REAL WORK.

Give your agent software it can use.

Connect through MCP or explore the Agent App catalog and choose only the capabilities your workflow needs.

EXPLORE AGENT APPS