What it is
Weave is the LLM side of Weights & Biases. Its SDK traces agents (conversations, turns, model calls and tool calls) and any Python or TypeScript function decorated with @weave.op, and it runs evaluations with LLM judges and custom scorers, including RAG retrieval checks. Integrations patch the OpenAI Agents SDK, the Claude Agent SDK and Google ADK automatically, and agents already instrumented with OpenTelemetry can send spans to Weave's OTLP endpoint without the SDK.
The SDK is Apache 2.0-licensed; version 0.53.13 was published on 22 October 2026. Traces go to the hosted W&B service, now sold as part of CoreWeave Forge after CoreWeave's 2025 acquisition of Weights & Biases. Forge's free plan includes 1 GB of Weave data ingestion a month; Pro starts at $60 a month with 1.5 GB, then $0.10 per MB. Dedicated deployments are Enterprise-only.
In September 2026 CoreWeave put Agent Lens into public preview and its docs call it the successor to Weave for agent tracing. Both read the same trace data, and Weave remains the home for evaluations, function-level traces and guardrails.
Our review
Weave earns its place inside an existing Weights & Biases account and struggles to justify itself outside one. Teams that already track training runs there can add @weave.op to an agent and see its traces and evaluations beside the models that produced them, and the automatic integrations for the OpenAI Agents SDK, Claude Agent SDK and Google ADK make the first trace a few lines of code.
The economics are the obstacle. The free plan includes 1 GB of Weave ingestion a month and Pro includes 1.5 GB, after which Pro charges $0.10 per megabyte and Enterprise lists $0.20. Agent traces carry full prompts and tool outputs, so a team logging 10 GB a month on Pro would pay about $930. Annual commitments bring volume discounts, but the list rate punishes verbose agents.
Hosting choices are narrower than the Apache 2.0 SDK suggests. The multi-tenant Forge service runs in CoreWeave's Google Cloud account in North America, holds a SOC 2 Type 2 report and is not HIPAA-compliant; HIPAA work needs Dedicated Cloud. Self-managing Weave means a licensed W&B Server deployment plus a ClickHouse cluster, so there is no free self-hosted path of the kind Langfuse or Arize Phoenix offer.
Direction is unsettled. CoreWeave's own docs call Agent Lens, in public preview since September 2026, the successor to Weave for agent tracing, while saying both already receive the same data and nothing needs migrating. Evaluations, function-level traces and guardrails stay in Weave, and the SDK still ships several releases a month, the latest on 22 October 2026. OpenTelemetry ingestion keeps the exit cost low, which is the strongest argument for trying it.
Where it earns its place
An ML platform team of eight that already logs fine-tuning runs to W&B
Compare a fine-tuned model against its base on the same agent evaluation
Decorate the agent with @weave.op, run a Weave evaluation with LLM-judge scorers for each model and view the results next to the training runs. Pro starts at $60 a month with up to 10 model seats; eval traffic beyond 1.5 GB a month bills at $0.10 per MB.
A backend engineer running an OpenAI Agents SDK support bot
Debug which tool call broke a conversation
The Weave integration patches the Agents SDK, so sessions, turns, model calls and tool calls appear in the Agents view without manual spans. Sampling or trimming logged payloads keeps ingestion inside the allowance.
A finance lead approving observability spend for a 20-engineer AI team
Forecast the monthly tracing bill before signing
The bill is driven by megabytes ingested, not seats: 10 GB a month on Pro is about $930. Ask for annual-commitment volume pricing, and price Dedicated Cloud separately if HIPAA or data residency applies.
Before you commit
- Setup
- pip install weave on Python 3.10 or newer (a TypeScript SDK also exists), sign in to a CoreWeave Forge W&B account, and call weave.init; the first trace takes minutes.
- Integrations
- Automatic patching for the OpenAI Agents SDK, Claude Agent SDK and Google ADK, an OTLP endpoint for any OpenTelemetry-instrumented agent, and the wider W&B platform for models, artifacts and reports.
- Data and privacy
- Traces are stored on CoreWeave Forge, hosted in Google Cloud North America regions, with a SOC 2 Type 2 report; Forge is not HIPAA-compliant, Dedicated Cloud is the option for that. SSO, SCIM and audit logging are documented for the platform. No published security advisories on the SDK.
- Cost at scale
- Free 1 GB a month; Pro from $60 a month with 1.5 GB, then $0.10 per MB; Enterprise overage $0.20 per MB. 10 GB a month on Pro is about $930. Academic licences get 25 GB a month and up to 100 seats free.
- Limits
- 189 open issues and pull requests; agent tracing is shifting to Agent Lens, still in preview; self-managed deployments must keep ClickHouse versions compatible with each W&B Server upgrade.
What's good
- Automatic tracing for OpenAI Agents SDK, Claude Agent SDK and Google ADK
- Traces and evals beside W&B training runs
- Accepts OpenTelemetry spans without the SDK
- Apache 2.0 SDK with several releases a month
What's not
- Ingestion overage of $0.10/MB on Pro, $0.20/MB on Enterprise
- Self-hosting needs a W&B Server licence and ClickHouse
- Multi-tenant cloud is North America only and not HIPAA-compliant
- Agent tracing being succeeded by Agent Lens
Best for
Teams already using Weights & Biases for model training who want LLM traces and evaluations in the same place.
Alternatives
W&B Weave: common questions
- Is W&B Weave worth it?
- Worth turning on if your models already live in Weights & Biases; at $0.10 a megabyte, a hard sell for anyone starting fresh. We rate it 3.6 out of 5.
- How much does W&B Weave cost?
- SDK free under Apache 2.0. Hosted: Free (1 GB Weave ingestion/month); Pro from $60/month (1.5 GB, then $0.10/MB); Enterprise custom ($0.20/MB overage listed), with dedicated deployment.
- What are the downsides of W&B Weave?
- Ingestion overage of $0.10/MB on Pro, $0.20/MB on Enterprise; self-hosting needs a W&B Server licence and ClickHouse; multi-tenant cloud is North America only and not HIPAA-compliant; agent tracing being succeeded by Agent Lens.
- Who is W&B Weave best for?
- Teams already using Weights & Biases for model training who want LLM traces and evaluations in the same place.
- How much does W&B Weave cost for a team or in production?
- Free 1 GB a month; Pro from $60 a month with 1.5 GB, then $0.10 per MB; Enterprise overage $0.20 per MB. 10 GB a month on Pro is about $930. Academic licences get 25 GB a month and up to 100 seats free.
- Is W&B Weave safe for company data?
- Traces are stored on CoreWeave Forge, hosted in Google Cloud North America regions, with a SOC 2 Type 2 report; Forge is not HIPAA-compliant, Dedicated Cloud is the option for that. SSO, SCIM and audit logging are documented for the platform. No published security advisories on the SDK.
- What are the best W&B Weave alternatives?
- The closest alternatives we have reviewed are Langfuse, Opik and Arize Phoenix.
Answers checked Oct 23, 2026.