What Is agav (2026)? Token-Saving Terminal AI Agent
By Antigravity Skills editorial team · Last updated
Short answer: agav is an open-source (Apache 2.0) coding agent that runs in your terminal and edits files, runs tests and checks its own changes. Its selling point is lower token spend: the README says it reuses cached context, stops unproductive repair loops, caps retries and isolates subagent exploration. It is an experimental beta, and its cost figures come from the authors.
Unofficial guide. Antigravity Skills is not affiliated with agav or Prapaa. Every fact below comes from the official GitHub repository, the authors' cost-efficiency write-up and the Show HN thread, read on November 18, 2026. The project is moving fast, so confirm commands and limits on the official pages before relying on them. We did not install or run agav.
What is agav?
According to its README, agav is a "terminal-native autonomous agent" for real repositories and projects. You start it with agav inside a project, give it a task, and it can inspect code, edit files, run tests and verify the result from the command line. If you have used Claude Code or Codex in a terminal, the idea is the same; the difference agav claims is how carefully it spends tokens.
Other features the README lists:
- Providers: Anthropic, OpenAI, OpenRouter, NVIDIA NIM, Gemini, Vertex AI and Ollama for local models, with
/modelto switch mid-session and/fastand/deepshortcuts. - Non-interactive mode:
agav runandagav --print, with per-tool permissions and optional JSON Schema output, so you can pipe a log into it. - Sessions and memory: resume, branch, name, search and export sessions; per-project memory that carries across sessions;
/compactto free context. - Skills and subagents: reusable instruction bundles, a marketplace of service agents, and fresh-context subagents for scoped work.
- Safety controls: encrypted API keys, per-action tool permissions, human approval for destructive actions, and
/permissionto switch between ask, auto-accept and deny-writes.
How agav saves tokens
The README names four mechanisms. Reading them as the authors wrote them, not as guesses:
| Mechanism (README) | What it means in practice | Is it measured separately? |
|---|---|---|
| Reuse cached context | Repeated parts of the prompt hit the provider's prompt cache, which is billed far below fresh input | The cost study reports a 97.0% cache-hit share of input tokens, but no cache-off run |
| Interrupt unproductive repair loops | The agent stops going round the same failing fix | No ablation reported |
| Bound retries and budgets | Hard caps on how many times and how much it spends | No ablation reported |
| Isolate subagent exploration | A subagent reads the repo in its own context so the main thread does not resend it | No ablation reported |
The cost write-up is candid about this last column: it says no ablation was run, so the savings cannot be credited to any single feature. Treat the list as the design intent and the cache-hit figure as the only measured piece.
What the authors' own numbers show
The authors ran Terminal-Bench 2.1 (89 tasks, 5 trials each, 445 trials) against a published Codex 0.144.1 submission, with both using the same model at maximum reasoning effort. These are vendor-reported figures, not independent results.
| Measure | agav | Codex |
|---|---|---|
| Pass rate | 269 of 445 trials (60.4%) | 341 of 445 trials (76.6%) |
| Total estimated model cost | $122.86 | $241.45 |
| Median cost per trial | 15.2 cents | not stated |
Total cost came out 49.1% lower, and agav was cheaper on 71 of the 89 tasks. The README itself says this is "a cost result, not a claim of higher accuracy", and the numbers back that up: agav passed fewer trials. Costs use inferred per-token rates (not an official rate card), cover model tokens only, and count timeouts as failures. The README also lists 62.0% (186 of 300) on SWE-bench-Live Lite and 84.7% on Terminal-Bench 2.1 accuracy, noting that the latter submission was still under review when we read it.
The practical reading: if your tasks are cheap to retry and token bills matter, the trade may be worth testing. If a failed run costs you an afternoon, a lower pass rate is a real cost that the dollar figure does not capture.
How to install and run agav
The README gives these install routes. It says no Node.js is required because agav ships as a self-contained binary.
macOS or Linux:
curl -fsSL https://agav.dev/install.sh | bashWindows PowerShell:
irm https://www.agav.dev/install.ps1 | iexWindows Command Prompt uses a downloaded install.cmd, which you run and then delete. Release binaries are on the GitHub releases page. Piping a script into your shell runs whatever the server sends, so read the script first or use the release page if you prefer.
Run examples from the README:
agavstarts the interactive agent.agav --provider ollama --model llama3.2uses a local model.agav --provider openrouter --model openrouter/freeuses OpenRouter; the README says OpenRouter and NVIDIA NIM offer free daily credits with no paid plan required.agav -P "..."oragav run "..."runs a single prompt, andagav updateupdates the binary.
Limits and caveats
- Experimental beta. The README warns to expect breaking changes, rough edges and rapid iteration.
- Sandboxing coverage. The README lists command sandboxing for macOS and Linux. In the Show HN thread, a maintainer said it uses macOS Seatbelt, Linux bubblewrap and a Windows restricted token, applied to tool execution rather than the whole terminal session. The README text we read mentions only macOS and Linux, so check Windows behaviour yourself.
- Voice input is offline and English only.
- Background jobs are opt-in.
- Accuracy. Lower cost came with a lower pass rate in the authors' own test.
- Future pricing. In the thread, a maintainer said they plan to resell tokens of top models, with pricing not final, and that the open-source version would keep working as is. Nothing about a paid plan is in the README.
What Hacker News said
The Show HN post, "an open-source terminal agent designed to waste fewer tokens", had a score of 26 when we checked and only a handful of comments. A reader asked for a short summary, another asked how it differs from OpenCode and whether it was vibe-coded. A maintainer replied that AI was used heavily but changes go through review, tests and security checks, and that agav focuses on cost efficiency with minimal accuracy tradeoff plus secure execution. One commenter asked what practice it uses to cut token use; that question had no reply in the thread we read. The discussion is small, so it is not evidence either way.
FAQ
Is agav free and open source?
The README says agav is Apache 2.0 open source with no lock-in, hidden telemetry or closed hosted runtime required. You still pay your model provider unless you use a free-credit route or a local model.
Does agav really save tokens?
The authors report 49.1% lower total estimated cost than Codex on Terminal-Bench 2.1, with 97.0% of input tokens served from cache. They also report a lower pass rate, and no independent replication was available when we checked.
Which models does agav support?
Anthropic, OpenAI, OpenRouter, NVIDIA NIM, Gemini, Vertex AI and Ollama, per the README.
Is agav production ready?
The README labels it an experimental beta and warns about breaking changes.
Related on Antigravity Skills
- DeepSeek harness: another open-source agent harness and how to run it
- OpenCode Go: the price, models and limits of OpenCode's coding-model plan, the tool agav was compared to in the HN thread
- Claude Code "don't ask again": the permission trade-off that agav's ask, auto-accept and deny-writes modes also touch
- OpenClaw: a self-hosted, open-source agent you run yourself
- All skills