DEVELOPER GUIDE
Integrate Athena,
step by verified step.
Athena is a self-hosted Python framework in private developer preview. There is no public API or hosted service yet — this guide describes exactly what approved pilot customers receive and run today.
1. Get access
Athena's source repository is private. Access starts with a pilot inquiry: tell us the workflow you want to improve and how you will measure it. We confirm technical fit, agree scope and licensing, and then grant repository access. No public downloads, checkout, or self-serve signup exist yet.
2. Install
Requirements: Python 3.11 or newer, git, and (optionally) API keys for the model providers you want to connect. The core framework itself is standard-library only and runs fully offline with its built-in mock provider.
git clone https://github.com/Kshot3000/Athena.git
cd Athena
python -m pip install -e .
3. Verify your installation
Before connecting any provider, run the offline scoreboard and the test suite. Both must pass on a clean checkout — if they don't, stop and tell us what you saw; a broken baseline is a bug report we want.
python -m athena eval # routing scoreboard (train + holdout)
python -m athena eval-semantic # semantic routing suite
python -m athena eval-adversarial
python -m unittest discover -s tests
4. Connect model providers
Providers are read from environment variables only — never commit keys to the repository. With no keys set, Athena uses an offline mock provider so the full workflow (recall, propose, evaluate, adopt) can be exercised safely.
# Set only the providers you use, in YOUR environment:
# ANTHROPIC_API_KEY / OPENAI_API_KEY / XAI_API_KEY
python -m athena demo "Route a support ticket"
Model availability depends on your provider account, current model names, and endpoint compatibility. The council pattern lets different models propose and verify, so a single-provider setup works but is not the strongest configuration.
5. Understand the safety model before you build on it
- Evidence-gated memory. A lesson enters Athena's memory only with evidence attached — a score, a test result, a measured outcome. Unsupported claims are not remembered as fact.
- Gated improvement. Playbook and routing changes are adopted only when they improve the training split without regressing any protected split (train/holdout across every suite in the gate).
- Constrained self-modification. The code lab accepts data-only candidates, statically screened — candidate code is never executed during evaluation.
- Human control. Core code changes, account actions, and anything touching money require human review and approval. Athena has no blanket permission over your production systems; external actions need integrations you build and approve.
6. Build your evaluation first
Athena improves against a scoreboard, so the highest-value integration work is writing evaluation cases from your own workflow: real tasks, human-labelled expected outcomes, split into train and holdout sets. Start narrow — one workflow, twenty honest cases — and expand only when the loop demonstrates gains on cases it has never seen.
Support
Pilot customers get support through Kshot Applications. Include your Athena version, the exact command you ran, and its full output — the same evidence standard Athena holds herself to.
← Back to Athena