- Category
- AI Agents
- Rank
- No. 3574Tools index
- Pricing
- Open Source
- Type
- AGENT
- Use case
- Coding
- Interfaces
- CLI
- GitHub
- 183 stars
- Latest release
- v0.2.8
- Date
About
Agav is a terminal-native AI coding agent that inspects repositories, edits files, runs tests and verifies changes from the command line. It reuses cached context, bounds retries and budgets, and isolates subagent exploration to reduce token spend. It works with Anthropic, OpenAI, OpenRouter, NVIDIA NIM, Gemini, Vertex AI and Ollama, and is labeled an experimental beta.
At a glance
Agav is an open-source (Apache 2.0) coding agent that runs in your terminal against a real repository: it reads code, edits files, runs tests and tracks a plan. You bring the model, from Anthropic, OpenAI, Gemini, Vertex AI, OpenRouter, NVIDIA NIM or a local Ollama install. It is an experimental beta, and the project warns of breaking changes.
Good fit
Developers who want a terminal coding agent that is not tied to one model vendor and who care about API spend. The project says it reuses cached context, bounds retries and budgets, and cuts off unproductive repair loops. agav run with optional JSON Schema output also suits scripted or CI jobs, and sessions can be resumed, branched and exported.
Consider something else when
Skip it if accuracy matters more than cost. In the project's own Terminal-Bench 2.1 cost study with the same model, Agav passed 60.4% of trials against 76.6% for Codex while spending 49.1% less. Also skip it if you need a stable tool for a team, since the beta warns of breaking changes and rough edges.
Start here
Install the self-contained binary with curl -fsSL https://agav.dev/install.sh | bash (PowerShell users run irm https://www.agav.dev/install.ps1 | iex). Get an API key; OpenRouter and NVIDIA NIM offer free daily credits. Then try a read-only task first: agav run "List the files in this repository and explain its structure. Do not modify anything."
Cost and requirements
The software is free and needs no Node.js. You pay for whatever model provider you use, or run local models through Ollama at no API cost. The project reports a 15.2 cent median estimated cost per Terminal-Bench trial at GPT-5.6 Luna rates, which is a benchmark estimate, not a price. Commands can run in a sandbox on macOS and Linux.
Limits to know
It is a beta at version 0.2.8 with 190 open issues. Benchmark figures come from the project: a 62.0% verified result on SWE-bench-Live Lite and an 84.7% Terminal-Bench 2.1 score whose submission was still under review. Safety controls include per-action permissions, approval for destructive actions and a deny-writes mode.
Still unclear
- The project does not say which model produced its 62.0% SWE-bench-Live Lite and 84.7% Terminal-Bench 2.1 scores, so it is unclear how much of that accuracy carries over to the model you choose.
- The project names macOS and Linux for sandboxed commands but does not say whether Windows users get a sandbox, which matters if you plan to let the agent run commands unattended there.
Sources checked 2026-11-17. Source-based assessment.
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.