W&B Weave
Weights & Biases toolkit and Python SDK for tracing and evaluating AI agents and LLM applications.
W&B's open-source SDK for agent tracing and evals; the trace server is not in the repo.
Weave is Weights & Biases' toolkit for tracing and evaluating AI agents. The repository is the Apache-2.0 client SDK; traces go to a W&B account, so it suits teams already on W&B rather than people who want a self-hosted server.
What it is
After weave.init(), supported SDKs such as OpenAI Agents SDK, Claude Agent SDK and Google ADK are autopatched and show conversations, turns, LLM calls and tool calls in an Agents tab. For your own loops, start_conversation, start_turn, start_llm and start_tool wrap them, and @weave.op traces plain functions into the Calls tab; it also runs evaluations.
Who it's for
It suits Python developers who already use Weights & Biases, or are happy to sign up for it, and want agent traces with almost no setup. It does not suit anyone needing the whole stack in their own data centre from this repo, because the README says the trace server lives in wandb/core.
Key features
One-call agent tracing
weave.init() autopatches OpenAI Agents SDK, Claude Agent SDK and Google ADK.
Agents tab
Conversations, turns, LLM calls and tool calls are grouped in a dedicated Agents view.
Custom agent context managers
start_conversation, start_turn, start_llm and start_tool wrap your own agent loop.
@weave.op function tracing
Decorate functions to see them in the Calls tab.
Evaluation framework
Weave evaluates agents and LLM applications; the README places its code in weave/flow.
W&B account integration
Traces land in your Weights & Biases project and can sit beside its experiment tracking.
Up and running in 3 steps
Our honest verdict
What's great
- Tracing the supported agent SDKs needs only weave.init(), per the README quickstart.
- The Agents tab models conversations and turns, not only a flat list of calls.
- The Apache-2.0 SDK covers Python 3.10 or higher, and the project also ships a TypeScript package.
- It connects to the Weights & Biases platform, which you may already use for model experiment tracking.
Where it falls short
- A Weights & Biases account and API key are required (WANDB_API_KEY); the README quickstart does not describe running your own trace server.
- The README says the trace server code is developed in wandb/core, so this repo is the client, not a self-hostable backend.
- The README says the Weave engine and Weave boards code is on pause while the team focuses on tracing and evaluations, and a cleanup is in progress.
- The quickstart example also needs an OpenAI API key, and the older start_session API now emits a DeprecationWarning.
Pricing & license
The SDK is free under Apache-2.0, but traces are stored in a W&B account. The pricing page (wandb.ai/site/pricing, now titled CoreWeave Forge Pricing) lists Free at $0 a month; Pro starting at $60 a month with a 30-day free trial; and Enterprise on custom plans. Weave ingestion is listed at 1 GB a month on Free and 1.5 GB on Pro, with extra ingestion billed per MB.
Open remote AI-training roles
767 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I self-host Weave from this repo?
The README does not describe it. It says the trace server Weave talks to is developed in wandb/core, so this repo is the client and evaluation code.
Do I need a W&B account?
Yes, the README lists a Weights & Biases account and API key as prerequisites, with a free tier available.
Which agent frameworks work with one line?
The README names OpenAI Agents SDK, Claude Agent SDK and Google ADK, for which weave.init() is enough.
