Quickstart
Jejoo analyzes AI-agent execution traces and suggests ways to improve cost, latency, and reliability. Start with a synthetic dataset, then try your own traces.
The CLI processes traces on your machine and retains the full trace files and canonical execution facts locally. It sends representative samples, filtered by your redaction settings, plus a schema/profile, metadata, and bounded evidence to the service. Optional workflow files are sent in full. See what stays local and what is shared.
1. Install
You need Python 3.11 or newer and an internet connection to the Jejoo service. The examples use a macOS/Linux shell.
python3 -m venv .venv-jejoo
source .venv-jejoo/bin/activate
python -m pip install --upgrade jejoo-cli
jejoo --version
On Windows PowerShell, activate with .venv-jejoo\Scripts\Activate.ps1. Use the download links below in place of curl if needed.
jejoo-cli is currently supplied by Scail as a prerelease wheel. Install it with python -m pip install /path/to/jejoo_cli-0.1.0-py3-none-any.whl instead of the PyPI command above. The PyPI install and upgrade commands will work after package publication.
No Jejoo account, API key, or project ID is required for anonymous use. On first use the CLI creates an installation identity and stores private credentials under ~/.config/jejoo. Anonymous usage is subject to service allowances.
If you previously installed optimaize-cli, uninstall it before installing jejoo-cli: the distributions share installed files. Existing optimaize commands remain compatibility aliases.
2. Get sample traces
Download sample-traces.jsonl. It contains three synthetic research runs, including a repeated search and model token/cost measurements. It contains no customer data, and its costs are illustrative.
curl -fL https://docs.scaillabs.ai/examples/sample-traces.jsonl -o sample-traces.jsonl
To use your own traces, see Input Formats.
3. Run an analysis
jejoo analyze sample-traces.jsonl \
--api-base-url https://app.scaillabs.ai \
--output-dir .jejoo/first-analysis
Both --api-base-url and --output-dir are required. Keep the CLI running while it prepares local facts, waits for normalization, and answers bounded evidence requests. It prints progress while the service investigates.
The service receives a schema/profile, representative redacted samples, workflow context, and bounded evidence. Review Privacy & Data Handling before using real traces.
4. Read the report
A successful run prints Analysis status: completed and a Report: path. Open:
.jejoo/first-analysis/analysis-report.md
The same directory contains recommendations.json for structured results and analysis-result.json for job status and the audit record. The CLI saves a local report; it does not currently print a shareable web-report URL.
Read the suggested changes, why they were proposed, what remains uncertain, and how to check whether they work. The sample may reveal repeated work or latency concentration; specific recommendations are not guaranteed. See Understanding Results.
Add .jejoo/ to your application's .gitignore. The output directory can contain sensitive local data.
Try your own agent
Replace sample-traces.jsonl with your export and choose a fresh output directory. Optionally include the source files that implement the agent and its retry or routing logic:
jejoo analyze traces.jsonl \
--workflow agent.py \
--workflow tools.py \
--objective 'Reduce latency while preserving answer quality.' \
--api-base-url https://app.scaillabs.ai \
--output-dir .jejoo/my-agent
Workflow files are optional. Their full contents and relative filenames are sent to the service and its models; trace sample redaction flags do not redact workflow files.
You can also retrieve traces directly from LangSmith or Langfuse.
If a run is interrupted
Rerun the same command with the same inputs and output directory to resume saved work. For changed inputs or a new analysis, use a fresh directory. Increase --wait-timeout-seconds if needed. See recovery and troubleshooting.
To update later, activate the same virtual environment and run python -m pip install --upgrade jejoo-cli.