Agav

Agav is an open-source, terminal-native AI coding agent (and harness) for macOS, Linux and Windows. Its pitch is that cost efficiency belongs in the harness: cacheable prefixes, bounded tool output, pruned searches and stop rules for unproductive loops — not only a cheaper model. On their Terminal-Bench 2.1 comparison with the same GPT-5.6 Luna rates, Agav's estimated spend was about 49% lower than Codex, with a lower pass rate (60.4% vs 76.6%).

Features include model switching (including Ollama), MCP tools, subagents, local project memory, SSO/policies for teams, and a ~27–39 MB install via curl | bash.

Harness levers

• Conversation-prefix caching and stable tails so git/time noise does not bust the cache.
• Tool output capped (~40 KB / 2k lines) with overflow spilled to disk.
• Search that skips node_modules / .git / dist.
• Iteration budgets and receipts so spend is inspectable.

Trade-off

Cheaper ≠ stronger: their published TB2.1 numbers show lower cost and lower pass rate versus Codex. Evaluate on your own tasks.

1. Install: curl -fsSL https://agav.dev/install.sh | bash.
2. Configure your model endpoint (cloud or Ollama) and optional MCP servers.
3. Run Agav in a repo; use /model to switch; inspect usage receipts after long jobs.
4. For teams, look at SSO / policy docs on agav.dev/teams.