Agav
Terminal-based coding agent for working on real repositories, with cost-aware model calls
- Pricing
- Open source
- Platforms
- CLI, macOS, Linux
- Listing updated
- today
Similar agents
About Agav
Agav is an open-source agent for coding. It is available on CLI, macOS, Linux, Windows, self-hosted and iOS.
The README says Agav runs as a single self-contained binary in the terminal, started with the agav command, and can be pointed at a range of model providers including Anthropic, OpenAI, OpenRouter, NVIDIA NIM, Gemini, Vertex AI and local Ollama models. Models can be switched mid-session, and the project states that API keys are stored encrypted.
The project emphasises how much each model call costs. It reuses cached context, caps retries and budgets, interrupts repair loops that make no progress, and isolates subagent exploration so the same repository context is not resent. Plans are tracked, sessions can be resumed, and destructive actions can require human approval, with per-tool permissions and sandboxed command execution on macOS and Linux.
It suits developers who work inside a repository from the command line and want an agent they can script in CI through agav run, with optional JSON Schema output. The README also lists an experimental beta status, so breaking changes are expected.
What it does
- Multi-provider model support
- Connects to Anthropic, OpenAI, OpenRouter, NVIDIA NIM, Gemini, Vertex AI and Ollama, and lets you switch models within a session.
- Cached, budgeted model calls
- Reuses cached context and applies limits on retries, budgets and repeated repair loops to reduce spend.
- Non-interactive CI mode
- The agav run command and print mode run scripted tasks with per-tool permissions and optional JSON Schema output.
- Resumable sessions and memory
- Sessions can be resumed, branched and exported, and cross-session memories are kept per project.
- Permission controls
- Switch between ask, auto-accept and deny-writes modes, with human approval available for destructive actions.
- Repository-aware editing
- Uses LSP queries, notebook support, test running, undo and a planning view when editing code.
- Offline voice input
- The /voice command dictates prompts using offline, English-only transcription.
Good for
- Exploring and explaining an unfamiliar codebase from the terminal
- Making scoped code changes with plans and approval steps
- Running an agent non-interactively inside CI pipelines
- Diagnosing errors by piping log output into a prompt
- Running coding tasks on local models through Ollama
Pricing
Open source
Apache 2.0 open source; model API usage depends on the chosen provider.
Agav FAQ
Is Agav open source?
The README says Agav is released under the Apache 2.0 licence.
Do I need a paid plan to start?
The README says OpenRouter and NVIDIA NIM offer free daily LLM API credits, with no paid plan required to use them.
Which operating systems support sandboxed commands?
The README says commands can run in a sandbox on macOS and Linux.
How does Agav use Terminal-Bench results?
The README reports an 84.7% accuracy result on Terminal-Bench 2.1 and a cost study comparing it with Codex, with the methodology published on the project site.
Is Agav free?
Yes. Agav is open source, so you can use and self-host it for free.
What are the best alternatives to Agav?
The closest alternatives to Agav are openclaude, devika, CoderBaba, munder-difflin and Devon.
Where this comes from
- Description: written from the project's README.


