Installation
TraceForge runs on Python 3.11, 3.12, and 3.13.
pip install traceforge-toolkit # or: uv add traceforge-toolkit
Everything ships with a single install, no extras to choose. The pipeline, enricher,
classification engine, risk scoring, live phase/boundary/title structuring, the
governance/assessment engine, all eight storage sinks, and the traceforge CLI are all
included.
The titler model weights
The activity/step titler model weights (~90 MB) live in a separate
traceforge-title-model package that traceforge depends on, so pip install traceforge-toolkit
pulls them automatically. The weights are hosted on PyPI (primary) and mirrored on this
repo's title-model-v* GitHub releases.
Because the weights are a hard dependency they install automatically. If a checkout left them as Git-LFS pointer stubs, or the install is otherwise broken, repair it with:
pip install --force-reinstall traceforge-title-model
To serve a relocated or custom titler, set TRACEFORGE_TITLE_MODEL to a directory containing the
encoder.onnx / decoder.onnx / tokenizer.json triad.
The phase and boundary models ship inside the base wheel, so only the larger titler is split out.
:::note Lightweight runtime
No GPU, no torch, and no transformers at runtime. The structuring models run on a light ML
stack that loads lazily, so an unused subsystem costs nothing. See
Live Structuring for the full runtime footprint.
:::
Optional sink dependencies
Only the S3 sink needs an extra third-party package. Install it with the s3 extra when you use
that sink:
| Sink | Extra | Install |
|---|---|---|
S3Sink | boto3 | pip install traceforge-toolkit[s3] |
Every other sink works out of the box. ParquetSink is built on pyarrow, which ships with the
base package, so no separate install is needed.
Verify the install
traceforge --help # top-level command group
traceforge detect # discover installed agent frameworks
Next: run your first pipeline in First Run.