Weights & Biases Weave

Weave traces and evaluates AI agents and LLM applications: automatic tracing via @weave.op decorators or OpenTelemetry-compatible SDKs, a flexible evaluation framework for scoring outputs against custom or LLM-judge scorers, a Playground for testing prompts/models against production traces, and Guardrails (pre-built scorers for toxicity, bias, PII, hallucination). As of mid-2026 W&B has repositioned Weave specifically around multi-agent, multi-turn systems — traces are organized into sessions/turns/steps/tools/sub-agents rather than flat call logs, which is a real structural difference from tools built for simpler single-call tracing. It also ships an MCP server so coding agents (the vendor names Claude Code) can read live production traces and run evaluation loops autonomously.

Free tier: $0/mo, up to 5 model seats, unlimited Weave seats, 5GB/mo storage, 1GB/mo Weave data ingestion (trace metadata + logged inputs/outputs), limited free inference credits. Pro: starts at $60/month billed monthly (30-day free trial), up to 10 model seats, 100GB/mo storage ($0.03/GB beyond that), 1.5GB/mo Weave ingestion ($0.10/MB beyond that — works out to roughly $100/GB, a steep step past the included allotment), plus a small $5/mo inference credit. Enterprise is custom-quoted. W&B explicitly caps Pro at companies under 50 employees and says larger customers "will be required to transition to Enterprise." Inference (running open-source hosted models through W&B) and the ARIA agent product are billed separately, per token, on top of any plan — full token-rate tables are published at wandb.ai/site/pricing/tokens. Storage is metered on a rolling 30-day GB-day average, and the vendor publishes the exact formula.

W&B offers a self-managed deployment ("W&B Server") that customers run on their own AWS/GCP/Azure account or on-prem, distributed via a Kubernetes operator and requiring a licensed connection back to deploy.wandb.ai. Required infrastructure includes Kubernetes, a MySQL 8.4.x database, S3-compatible object storage, and Redis — this is real infrastructure investment, not a Docker-compose weekend project. A "Personal" free tier (wandb server start) exists for non-corporate individual use only. W&B's own docs recommend the fully managed Dedicated Cloud over self-managed for most customers, and steer regulatory-sensitive buyers there for SOC 2/HIPAA coverage rather than self-hosted.

Weights & Biases was acquired by CoreWeave (Nasdaq: CRWV), a GPU-cloud infrastructure company, with the deal completed May 5, 2025 (CoreWeave press release, PR Newswire). Reported deal value in press coverage at the time was roughly $1.7B (Crunchbase). This matters for continuity risk in the opposite direction from most of our smaller listings: W&B is not an independent startup that might run out of runway, but it is now a product line inside a much larger infrastructure company whose core business is GPU compute, not developer tooling — its trust/subprocessor documentation for Weave now routes through CoreWeave's own Trust Center, and CoreWeave is itself listed as a subprocessor for "cloud infrastructure administration."

W&B publishes SOC 2 Type 2, ISO/IEC 27001:2022, 27017:2015, and 27018:2019 certifications, and states HIPAA compliance and NIST 800-53 alignment on its security page. It runs a public bug bounty (scoped only to qa.wandb.ai / api.qa.wandb.ai, tiered $50-$1,000 payouts) and lists its subprocessors, including AWS/Azure/GCP, ClickHouse, Datadog, and — notably — Anthropic and OpenAI as "Generative AI services" subprocessors (relevant if you're piping customer prompts through Weave's LLM-judge scorers). One HIGH-severity CVE (CVE-2024-7340, path traversal/arbitrary file leak, CVSS 8.8) was reported against the weave PyPI package; it was fixed in version 0.50.8. Current release is 0.53.6 (checked 2026-08-27), so the fix has been in place for many releases. No open GitHub security advisories are currently listed for wandb/weave.

wandb/weave on GitHub: 1,121 stars, 164 forks, Apache-2.0 licensed, last push 2026-08-26 (checked same day) — actively maintained, with 229 open issues and continuous PR merges from the W&B team. PyPI package weave is at version 0.53.6, uploaded under active release cadence; the README itself notes the repo is mid-cleanup, having deprioritized older "Weave engine"/"Weave boards" code to focus specifically on tracing and evaluations — worth knowing if you find outdated references to those features in older docs or forum posts.

No independent, verifiable review-site rating (G2, Capterra) was accessible for this run — G2's Weights & Biases review page blocked both direct fetch and proxy access. We could not find third-party benchmark comparisons of Weave against competing LLM-observability tools (Langfuse, Arize Phoenix, LangSmith) with concrete performance or cost data; any such comparison would currently rely on vendor-authored content. A public, standalone status/uptime page for the Weave/W&B service (as distinct from CoreWeave's own infra status page) was not found.