Agentagon 0.1.0¶

Agentagon brings evidence-backed audits and measured improvements to Codex and Claude Code. The CLI captures code and trace evidence, prepares evaluations, compares fixes against frozen baselines, and serves a local dashboard. The ag plugin supplies seven coding-agent workflows.

Install¶

Requires Python 3.12+ on macOS or Linux, and Git for evaluation/fix workflows.

pipx install agentagon
agentagon install --host codex
# Or: agentagon install --host claude-code

Restart the coding host after registration. Wheels and complete source archives are available on PyPI and this release; SHA256SUMS records their SHA-256 hashes. The source archive also builds the website and documentation.

Compatibility and privacy¶

Core local workflows require no Agentagon account. Your coding host supplies model access. Optional Intelligence and remote runners require separate setup. Anonymous usage telemetry is disabled by default; enable it explicitly with agentagon _internal setup --scope user --set telemetry.enabled true. An enabled setting can still be overridden with AGENTAGON_TELEMETRY_DISABLED=1. See the documentation for the exact fields.

Material for MkDocs remains the documentation theme. Required license notices are retained in the packages and site files; visible branding is Agentagon.

Validation and limits¶

Local release validation on September 10, 2026, with Python 3.13 on macOS: 707 tests passed, including the eight dashboard browser tests; nine optional integration tests were skipped. Python lint/format checks, JavaScript and shell syntax, Actionlint, secret scans and the resolved dependency audit passed. The wheel's resources and synthetic quickstart worked in a fresh environment outside the checkout. The complete website built from the source archive; five site checks passed and the installation page was checked in Chromium.

The release workflow requires passing tests, browser tests, source-package/site checks, secret scanning, dependency auditing, and CodeQL on the tagged revision of protected main. It validates package metadata and exercises the installed wheel and synthetic quickstart outside the checkout before publishing.

Native-host registration and live SSH/E2B integrations are optional checks and are not established by the regular CI suite. Model evaluations require explicit access and execution limits; a prepared evidence packet is not a passed audit.