Origami Trace

origami-trace ships the public otrace client for recording experiment runs in a hosted Trace project and reading those hosted runs back from the command line or MCP.

Use it when a training, evaluation, or data-processing run should upload its command, setup, logs, metrics, tables, and artifacts so the results are inspectable later.

Install

pip install origami-trace

Authenticate

Create a Trace project API key in the Trace dashboard, then initialize a workspace:

export TRACE_API_KEY="trc_your_key_here"
otrace init
otrace whoami

Run An Experiment

otrace run \
  --intent "Evaluate baseline on validation split" \
  --setup "Baseline config with validation data" \
  --expect-artifact outputs/metrics.json \
  --expect-log outputs/eval.log \
  -- python eval.py --config configs/baseline.yaml

If the project does not already write queryable outputs, create starter output files:

otrace scaffold outputs

Public Commands

otrace init
otrace login
otrace whoami
otrace scaffold outputs
otrace run
otrace import
otrace runs
otrace show
otrace artifacts
otrace download
otrace mcp

The published package is intentionally a client: it captures/uploads run evidence and queries/downloads hosted run data through the Trace API. The hosted backend and app implementation are not part of this package.

Query Hosted Runs

otrace runs --query baseline
otrace show run_123 --json
otrace artifacts run_123
otrace download artifact_123 --output outputs/metrics.json

MCP

Run the stdio MCP server:

otrace mcp

The public MCP tools expose hosted run listing, run detail lookup, artifact listing/download, and RunSpec validation.