A Python coding agent built for measured iteration.
Ava is a compact Python coding agent with a CLI, local Web UI, and embeddable API. It reads and edits code, runs commands, and resumes work from saved sessions. Harness changes are measured against independently graded coding tasks.
Quick start · Benchmark · Usage and configuration · Architecture
Requires Python 3.12+, uv, and macOS or Linux.
git clone https://github.com/SmartAI/ava.git
cd ava
uv sync
export ANTHROPIC_API_KEY=...
uv run ava --serveOpen http://127.0.0.1:8777 and choose a project, or use the CLI:
uv run ava -p "Explain this project"
uv run ava -c -p "Add tests for the main failure paths"For OpenAI, set OPENAI_API_KEY and add --provider openai. For an existing
Codex CLI login, use --provider codex. Select a supported model and reasoning
level with --model and --effort.
Ava is alpha software. Tools run with your local permissions and without per-call approval; session logs may contain sensitive content.
For the optional Qt Quick desktop app, run uv run --extra desktop ava-desktop --project ..
See the desktop usage guide for setup and current scope.
On September 8, 2026, Ava and Pi 0.85.1 completed 22 SWE-bench Pro development tasks across 11 repositories, with one attempt per agent per task.
Both used gpt-6-astra with medium reasoning through Codex OAuth.
| Measure | Ava | Pi |
|---|---|---|
| Tasks solved | 11/22 (50.0%) | 10/22 (45.5%) |
| Input tokens, including cached input | 4,571,720 | 4,275,286 |
| Output tokens, including reasoning | 87,128 | 81,101 |
| Tool calls | 515 | 475 |
| Mean agent execution, minutes/task | 3.65 | 3.18 |
Both agents solved 10 tasks; Ava alone solved one. All 44 attempts were graded, including one interrupted Pi attempt. This small, equally weighted repository sample does not establish general superiority or a full 731-task benchmark score. Per-request subscription cost was not measured.
Evaluation uses fresh environments, matched model settings and budgets, reference and unchanged-workspace controls, and independent patch grading. Failed attempts are retained in scores and resource totals. Detailed experiment records stay local; CI does not run evaluations.
Benchmark results and limitations · How to run evaluations
- Usage and configuration: CLI, Python API, custom providers and development.
- Architecture: agent loop, tools, context and session recovery.
- Evaluation guide: setup, controls, comparisons and session diagnosis.
- Port notes: differences from the original C++ runtime.
MIT © 2026 Min Liu.