Weave

license

Apache-2.0(OSI, SDK/client repo wandb/weave)

source

partial(the Apache-2.0 repo publishes the SDK and a ClickHouse-backed trace-server library, but no web UI, no deployment manifest and not the HTTP service that fronts the trace server, which its own trace_server/README.md places in W&B's closed core repository)

backend

commercial(W&B SaaS platform hosts traces/evals

eval+tracing-client

OSS

Only part of the product is published, which puts this on the source-available rung at 2 above any license test. wandb/weave is Apache-2.0 and publishes more than an SDK - weave/trace_server/ holds a working ClickHouse implementation with schema, migrations, query builders, workers, Kafka, file storage and OpenTelemetry ingest - but the repository carries no web UI, no docker-compose or Helm chart, and not the HTTP service that fronts the trace server. The repository's own weave/trace_server/README.md diagrams that service as living in core, which is W&B's closed repository; the package README's stated prerequisite is a Weights & Biases account; and W&B's deployment docs say a self-managed install needs a W&B Server license obtained before installation. You cannot stand the product up from what is published, which is what partial source means here.

• https://github.com/wandb/weave recorded 2026-06-04

  Apache-2.0 license on the Weave SDK repo; LLM logging/tracing/eval toolkit

• https://api.github.com/repos/wandb/weave/contents/ recorded 2026-08-12

  Repository root, listed to establish what ships. LICENSE (Apache-2.0 per the license endpoint), Makefile, noxfile.py, pyproject.toml, bin, dev_docs, examples, scripts, sdks, tests, trace_server_mock, weave. There is no docker-compose, no chart, no deploy or infra directory, and no web/ or ui/ - so no published way to run the application. sdks/ contains only `node`. bin/ contains only codex_setup.sh and run_setup_in_codex.sh. The server-side code that does ship is the library under weave/trace_server/, which includes clickhouse_trace_server_batched.py, clickhouse_schema.py, migrations, workers, kafka.py, file_storage.py and an opentelemetry package.

• https://raw.githubusercontent.com/wandb/weave/master/weave/trace_server/README.md recorded 2026-08-12

  The repository's own architecture diagram, which names the piece that is missing. The call-start sequence runs UserCode -> OpExecution -> GraphClientTrace -> RemoteHTTPTraceServer -> "(in core) trace_server.py TraceWebServer <FlaskApp>" -> ClickHouseTraceServer -> ClickHouseDB. The web service tier is annotated as living in core, W&B's closed repository, not in this one. The published RemoteHTTPTraceServer is the client that posts to it.

• https://raw.githubusercontent.com/wandb/weave/master/README.md recorded 2026-08-12

  The vendor's own prerequisite for using the published package, stated plainly under Prerequisites: "Python 3.10 or higher" and "A Weights & Biases account (free tier available)". The quick start is `pip install weave` then `weave.init("my-project-name")`, which registers against that account. Nothing in the README describes running Weave without W&B.

• https://docs.wandb.ai/guides/hosting/ recorded 2026-08-12

  W&B's deployment options: Multi-tenant Cloud, Dedicated Cloud and Self-Managed. On licensing it states "Some features and functionality require an Enterprise license" and routes readers to an Enterprise trial. The self-managed page adds "A W&B Server license authorizes your deployment. You must obtain a license before installation." So the on-premises route runs a licensed W&B Server distribution rather than a build of wandb/weave.

The weave package draws 1,052,916 PyPI downloads a month, just inside the 1M-10M band; it sits close to the boundary, and a further fall would put it at 3. It rides the established Weights & Biases install base, with ML teams adding LLM observability without leaving W&B.

• https://pypistats.org/packages/weave recorded 2026-08-13

  Downloads last month 1,052,916 for the weave package (last week 269,722)

A competent end-to-end tracing and evaluation suite: auto-logged OpenAI and Anthropic calls with inputs, outputs, tokens and cost; LLM-as-judge and code scorers running experiments against datasets; datasets and annotation; cost tracking; and framework integrations. Dedicated prompt management and versioning are lighter than MLflow's or Langfuse's, so this sits at the anchor tier of 4 rather than above it. The README frames the product as logging and debugging model inputs, outputs and traces plus rigorous evaluations, and says Weave boards are on pause "as we focus on Tracing and Evaluations", which narrows the surface rather than widening it: the band holds at Langfuse's 4, but with less headroom than the rest of this note implies.

• https://raw.githubusercontent.com/wandb/weave/master/README.md recorded 2026-08-13

  README - use Weave to "Log and debug language model inputs, outputs, and traces", "Build rigorous, apples-to-apples evaluations for language model use cases" and "Organize all the information generated across the LLM workflow, from experimentation to evaluations to production". It also records that "Weave boards" are on pause "as we focus on Tracing and Evaluations".

• https://github.com/wandb/weave recorded 2026-06-04

  logging/debugging LLM traces + rigorous evaluations feature set