How is this different from connecting Weights & Biases MCP directly?

Connected directly, an agent sees every tool the server exposes and holds the credential itself. Through TrustGate you choose which tools each agent gets, and every call is logged against the agent that made it.

What if the Weights & Biases MCP API changes?

We track the Weights & Biases MCP schema, so the tool catalogue stays current. Tools you have already allowed keep working, and anything new stays denied until you enable it.

Can I give different agents different Weights & Biases tools?

Yes. Access is set per agent: one can get a handful of read tools, another the full set. You grant the whole MCP or single tools within it.

Where can I see what my agents did in Weights & Biases?

Every call is logged against the agent that made it, with the tool it used. Activity can be sent on to your own monitoring.

Do I need to run the Weights & Biases MCP server myself?

No. TrustGate connects to the Weights & Biases MCP server and sits in front of it. Your agent only talks to your TrustGate URL, never to Weights & Biases directly.

Connect Weights & Biases MCP to your AI agents, securely.

Q: Is Weights & Biases MCP secure?

The server itself exposes 34 tools, and some of them can write to your Weights & Biases account. Behind TrustGate an agent only sees the tools you allow, requests and responses are inspected, and every call is logged. Anything outside the allowlist is denied before it reaches Weights & Biases.

One secured endpoint for every Weights & Biases call your AI agents make, scoped per user, inspected, and logged.

Get started for free

Get a demo

Try Weights & Biases MCP now

What your agents can ask once Weights & Biases MCP is connected.

Select the tools your AI agents can call

34 tools from the Weights & Biases MCP schema.

Showing 34 tools

infer_trace_schema_toolreadDiscovers the field names, types and sample values present in traces. W&B recommends calling this first so later queries can name precise columns.
query_weave_traces_toolreadQueries LLM traces with a detail_level of schema for structural fields only, summary for truncated inputs and outputs, or full for everything untruncated, which means every logged prompt and completion.
count_weave_traces_toolreadCounts traces matching a filter and returns storage metrics, without returning the trace contents.
resolve_trace_roots_toolreadResolves individual spans back to the root traces they belong to.
list_weave_agents_toolreadLists agents with aggregated statistics for invocations, tokens, duration and errors.
list_weave_agent_versions_toolreadReturns per-version statistics for one agent, for spotting a regression between versions.
query_weave_agent_spans_toolreadQueries individual agent, LLM and tool spans, filterable by agent, model and time.
get_weave_agent_span_stats_toolreadReturns time-bucketed metric series for tokens, cost, latency and error rate.
list_weave_agent_custom_attributes_toolreadDiscovers the custom attribute keys present on agent spans.
search_weave_agents_toolreadFull-text and structured message search across agent conversations, grouped by conversation. The results are the actual message contents.
get_weave_agent_trace_toolreadReturns a structured chat or trajectory view for one trace, meaning a single turn.
get_weave_agent_conversation_toolreadReturns the multi-turn chat view for a whole conversation.
probe_project_toolreadDiscovers a project's useful metric and configuration fields with bounded samples.
get_run_history_toolreadReturns bounded time-series history with sparse multi-key outer-union sampling, joining requested metric series by step. It uses a fixed application-owned projection and accepts no caller-supplied GraphQL.
compare_runs_toolreadCompares selected metrics and configuration across several runs.
diagnose_run_toolreadDiagnoses a run using bounded metadata and history reads, for questions such as why a run diverged.
summarize_evaluation_toolreadSummarizes bounded evaluation results together with coverage metadata describing what the summary did and did not see.
list_registries_toolreadLists the model registries in an organization.
list_registry_collections_toolreadLists the collections within a registry, meaning which models it holds.
list_artifact_versions_toolreadLists the versions of an artifact collection.
get_artifact_details_toolreadReturns the full details of an artifact version.
compare_artifact_versions_toolreadDiffs two artifact versions against each other.
query_wandb_toolreadRuns structured W&B reads using documented APIs, or a bounded query-only GraphQL document for shapes the structured interface cannot express. Mutations, subscriptions, multiple operations and mixed-mode inputs are rejected before any backend work. It is read-only in every access mode.
query_wandb_graphql_toolreadA separate GraphQL tool retained local-only in the compatibility profile. It is not required to use GraphQL, which is available through query_wandb_tool in every supported workload.
list_entities_toolreadLists the entities, meaning teams, that the current API key can access.
query_wandb_entity_projectsreadLists the projects belonging to an entity.
create_wandb_report_toolwriteCreates a W&B report with markdown, headings, line plots, bar plots, run comparisons, custom Vega charts and ordered panel layouts. Omitted entirely in read-only mode.
log_analysis_to_wandbwriteLogs analysis metrics to W&B as a run, which then appears in the project alongside real training runs. Omitted entirely in read-only mode.
list_wandb_automations_toolreadLists the W&B Automations that alert on run metrics or status.
list_wandb_integrations_toolreadLists the registered integrations automations can target, such as Slack channels or webhooks.
aria_send_messagehigh riskStarts or continues asynchronous work with the hosted W&B agent, returning a turn handle rather than a result. It requires the explicit ARIA profile and an approved HTTPS origin, is local-only in managed deployments, and read-only mode omits it.
aria_get_turnreadPolls one ARIA turn for progress or its final result, for up to thirty seconds. Poll results are compact unless a bounded raw snapshot is requested.
aria_get_turnsreadPolls up to twenty outstanding ARIA turns concurrently within one shared polling window.
search_wandb_docs_toolreadSearches the official W&B documentation, proxied from docs.wandb.ai. It can be disabled with an environment variable when the documentation server is connected separately.

What Weights & Biases MCP exposes

Weights & Biases MCP groups its 34 tools into 9 categories, each with a suggested starting policy.

All 34 tools as JSON, ready to paste.

Category	Covers	Access	Suggested policy
Weave Agents	The OpenTelemetry agent-spans data plane: listing agents with aggregated invocations, tokens, duration and errors, per-version statistics, querying individual agent, LLM and tool spans, time-bucketed metric series for tokens, cost, latency and error rate, discovering custom attribute keys, full-text message search grouped by conversation, and structured trace and multi-turn conversation views	Read	Read-only, but these return the actual content of agent conversations, which for most organizations means customer inputs and model outputs. Present only in the explicit agents profiles; dedicated deployments reject all of them
Runs & Experiments	Discovering a project's useful fields and bounded samples, bounded time-series run history with sparse multi-key outer-union sampling, comparing metrics and configuration across runs, diagnosing a run from bounded metadata and history, and summarizing evaluation results with coverage metadata	Read	Read-only and bounded by design. get_run_history_tool uses a fixed application-owned projection that accepts no caller-supplied GraphQL, which is the right shape for a hot path and makes it safe to allow broadly
Artifacts & Registries	Listing model registries in an organization, the collections within a registry, the versions of an artifact collection, reading the full details of an artifact version, and diffing two versions	Read	Read-only across the model supply chain. The registry listing describes which models an organization has in production, so treat the output as internal even though nothing here promotes or changes an artifact
Weave Traces	Discovering trace field names, types and sample values, querying LLM traces with a detail level of schema, summary or full, counting traces with storage metrics, and resolving spans back to their root traces	Read	Call the schema tool first and query with precise columns. detail_level full returns untruncated inputs and outputs, which is every prompt and completion the application logged, so default to summary and treat full as a sensitive read
Query & Discovery	Structured W&B reads or a bounded query-only GraphQL document, a separate local-only GraphQL tool in the compatibility profile, listing the entities the API key can reach, and listing an entity's projects	Read	One tool name covers the whole readable W&B surface through caller-supplied GraphQL, so log the query document rather than the tool name. Mutations, subscriptions, multiple operations and mixed-mode inputs are rejected before any backend work
ARIA	Starting or continuing asynchronous work with the hosted W&B agent, polling one turn for up to thirty seconds, and polling up to twenty outstanding turns concurrently within one shared window	Read / Write	aria_send_message hands work to an autonomous hosted agent, so what it does next is not visible in the tool call. It requires the explicit ARIA profile plus an approved HTTPS origin, is local-only, and read-only mode omits it while keeping the two polling tools
Reporting	Creating W&B reports with markdown, headings, line plots, bar plots, run comparisons, custom Vega charts and ordered panel layouts, and logging analysis metrics back to W&B as a run	Write	These are the only writes in the base profiles and both are omitted entirely by read-only mode. A report is visible to the team, and log_analysis_to_wandb adds a run that will appear in other people's project views
Automations	Listing W&B Automations that alert on run metrics or status, and listing the integrations those automations can target, such as Slack channels or webhooks	Read	Read-only. Nothing here creates or modifies an automation, which keeps an agent from quietly adding or silencing an alert, and the integrations listing shows where existing alerts are routed
Documentation	Searching the official W&B documentation, proxied from docs.wandb.ai so data tools and documentation search arrive on one connection	Read	Harmless and worth allowing freely. It can be switched off with an environment variable when the documentation server is connected separately, and upstream errors never echo the query back

Connect Weights & Biases MCP to your AI agent

Four steps, under ten minutes.

01
Sign in to NeuralTrust
Open the NeuralTrust platform, sign in and go to Getting started, then Connect your tools.
02
Choose your agent
When asked which agent you use, pick Claude Code, Codex, Cursor, Windsurf, Copilot or your own code.
03
Connect Weights & Biases
Find Weights & Biases in the tool list and connect it.
04
Connect your agent
Follow the steps on screen to add the config to your agent, then send the test prompt.

Permissions down to the single tool

You do not allow Weights & Biases MCP or block it. You allow the exact tools each agent can call, one by one.

Weights & Biases MCP	34 TOOLS AVAILABLE
Code review agent	4 tools, read only
Issue triage bot	9 tools
Platform team assistant	all 34 tools

Tool by toolConnecting the server is not the unit of access. Each tool is granted on its own.
Default denyA tool you have not enabled is never exposed to the agent.
Approval on writesTools that change or delete things can require a human to confirm.

Is Weights & Biases MCP secure?

One endpoint instead of raw credentials. Every call is checked against policy and logged.

AI agentsConnect through your MCP client configuration.	NeuralTrustApply controls based on your configuration.	Weights & Biases MCP34 named tools in version 0.4. Which appear depends on three independent settings: a named tool profile, a workload of shared, dedicated or local, and an access mode of read-only or read-write. Recommended presets range from 17 tools for a dedicated models-only deployment to 39 for hosted multi-tenant with Agent Lens, whose additional tools are documented separately. Dedicated deployments reject the ARIA and agent-span tools, and endpoint configuration alone never enables a tool family.
Coding assistants	Connection & credentialsper user	Weave Agents8 tools
Internal copilots	Tool access policiesdeny by default	Runs & Experiments5 tools
Workflow automations	Request & response inspectioninline	Artifacts & Registries5 tools
Compatible MCP clients	Activity monitoringalerts and logs	All other tools16 tools