<div align="center">

Agav

A terminal-native AI coding assistant for real repositories

<p> <img alt="Version" src="https://img.shields.io/github/package-json/v/prapaa-ai/agav?style=for-the-badge&amp;label=version&amp;color=111"> <img alt="License" src="https://img.shields.io/badge/license-Apache%202.0-111?style=for-the-badge"> <a href="https://github.com/harbor-framework/terminal-bench-2-1/pull/225"><img alt="Terminal-Bench 2.1" src="https://img.shields.io/badge/Terminal--Bench_2.1-84.7%25_%7C_top_of_the_board-111?style=for-the-badge"></a> </p> </div> <div align="center"> <img src="https://www.agav.dev/preview.gif" alt="Agav preview" width="100%" style="border-radius: 16px;" /> </div>

Vision

• The best agentic coding harness in the world — not a wrapper around an API, but a complete autonomous agent that reads, reasons, edits, and verifies across real codebases.
• Terminal-first, no compromises — the terminal is where work happens; Agav meets you there with the full power of an IDE and the speed of the command line.
• An agent for everyone — equally useful to the senior engineer debugging a distributed system and the non-technical founder drafting a pitch deck.

What it does

Agav reads, searches and edits the repository you run it in, and runs the commands you'd otherwise run yourself. On Terminal-Bench 2.1 — all 89 tasks, five trials each, judge-audited trajectories — Agav scored 84.7% (377 of 445 trials, ± 0.84%), ahead of every entry on the current public board (submission in review).

• Seven providers — Anthropic, OpenAI, OpenRouter, NVIDIA NIM, Gemini, Vertex AI and Ollama, switchable mid-session with /model.
• Sandboxed commands — shell tools run under Seatbelt on macOS and Bubblewrap on Linux where either is available.
• Non-interactive mode — agav run and agav --print make the same agent scriptable from CI, with per-tool permissions and optional JSON Schema output.
• Sessions that survive — resume, branch, name, search and export past conversations; /compact reclaims context without starting over. Plans are saved per-session and picked back up on resume.
• Extensible — MCP servers, plugins, a skill marketplace and installable agents; delegate scoped work to fresh-context subagents.
• Lights-off operation — schedule tasks (/schedule), loop prompts (/loop) and watch files (/watch); specs go in, verified work comes out.
• Repository-aware editing — LSP-backed queries, notebook support, test running, and /undo for the last file change.

Tools

Agav ships with 19 built-in tools the agent calls on your behalf — reading, writing, searching, running commands, and talking to external services.

| Tool | What it does | | --- | --- | | read_file | Read files — text with line ranges, PDF/Office with page ranges, images as compressed visual previews | | write_file | Create or overwrite files; creates parent directories as needed | | edit_file | Surgical string replacement — find an exact string, replace its first occurrence | | run_command | Execute shell commands, sandboxed via Seatbelt, Bubblewrap or Docker | | grep_search | Recursive regex search across files, with optional file-glob filters | | find_files | Glob-based file discovery | | list_directory | List a directory's contents with file types and sizes | | web_search | Search the web; returns titles, URLs and snippets | | fetch_url | HTTP requests (GET/POST/PUT/DELETE/PATCH) with custom headers | | lsp_query | Language Server Protocol queries — diagnostics, definitions, references, hover | | read_notebook | Read Jupyter notebook cells with their outputs | | edit_notebook | Edit Jupyter notebook cells by index | | github | GitHub CLI integration — create and view PRs and issues | | overview | Codebase structure map showing the file tree and key symbols per file | | run_tests | Auto-detecting test runner (pytest, vitest, jest, go test, cargo test) with structured pass/fail results | | update_plan | Mark plan steps as in-progress, done or failed | | save_memory | Persist cross-session memories (user, feedback, project, reference) | | subagent | Spawn independent parallel subagents, each with their own context and tools | | activate_skill | Run a registered skill by name |

Skills

Agav ships with a set of built-in skills — reusable instruction bundles the agent can activate on its own or that you can trigger manually. Skills load in order from bundled → global (~/.agav/skills/) → project (.agav/skills/), with later entries overriding earlier ones.

| Skill | What it does | Trigger | | --- | --- | --- | | code-review | Review code changes for bugs, security issues, and improvements | auto + manual | | deep-research | Multi-source research on a topic with citations | manual | | diagnose | Diagnose and fix errors and bugs | auto + manual | | doc-gen | Generate documentation for code | auto + manual | | explain | Explain code in plain language | auto + manual | | git-commit | Generate a commit message from staged changes | auto + manual | | refactor | Suggest and apply code refactoring | auto + manual | | security-scan | Check code for security vulnerabilities | manual | | simplify | Reduce complexity and simplify code | auto + manual | | test-writer | Generate unit tests for existing code | auto + manual |

Browse and install additional skills from the marketplace with /skills, or drop your own into the skills directory.

Agents

Agav can delegate work to standalone agents — in-process or external — that carry their own tools, model preferences, and permissions.

• Native agents — JS/TS agents defined by an AGENT.md file with YAML frontmatter, running in-process with custom tools, model/effort overrides, MCP servers, and tool permissions.
• A2A agents — external processes that communicate over HTTP via the Agent-to-Agent protocol.
• Marketplace — install agents from git repos with /agents; repos are sparse-cloned, validated, and sandboxed.
• Origins — agents load from bundled → global (~/.agav/agents/) → project-local, the same cascade as skills.
• Creation — /agents → Create opens a wizard that builds an agent definition with an LLM-generated system prompt, workspace MCP server selection, and credential management.

Memory

Agav remembers things across sessions. Memories are scoped per project (identified by the git root hash) and stored as markdown files.

Four memory types:

| Type | What it holds | Examples | | --- | --- | --- | | user | Role, preferences, expertise | "I'm a data scientist", "prefer tabs" | | feedback | Corrections and confirmations | "don't do X", "yes, that approach works" | | project | Project decisions, deadlines, context | "we use PostgreSQL", "deadline is Friday" | | reference | Pointers to external resources | Linear boards, Slack channels, dashboards |

The agent saves memories proactively when it detects relevant information during a session. Manage them yourself with:

• /memory — list and manage saved memories
• /remember — save a memory manually
• /forget — delete a memory by name

Memories are automatically loaded into future sessions for the same project.

Planning

Agav creates multi-step plans for complex tasks and tracks progress visually. Plans are saved per-session and picked back up on resume.

• The agent creates plans automatically when a task has enough moving parts to warrant one.
• Each step carries a status: in_progress, done, or failed.
• /plan shows the active plan; /plan list, /plan <n> <status>, and /plan clear manage it.
• Ctrl+G toggles the plan detail panel.
• Plans survive /compact operations and session resumes — context gets reclaimed, the plan stays.

Quickstart

Run it inside a repository:

agav

Pick a provider and model, or take the defaults:

agav --provider openai --model gpt-4o
agav --provider openrouter --model openrouter/auto
agav --provider vertex-ai --model vertex/gemini-3.5-flash
agav -r                                  # resume a session (lists them if no id)

Non-interactive, for scripts and CI:

agav run "review the code in src/"
# Read-only audit: block every tool that isn't explicitly allowed
agav run --permission '{"*":"deny","read_file":"allow","grep_search":"allow"}' "audit dependencies"
# Deny one tool; the rest still run without confirmation
agav run --permission '{"write_file":"deny"}' "check for security issues"
agav -P "what does this project do?"
agav -P --stream "explain this repository"
cat error.log | agav -P "explain this error"

Keep it current:

agav update

Options

| Flag | Meaning | | --- | --- | | --provider, -p | anthropic, openai, openrouter, nvidia, gemini, vertex-ai or ollama (default: anthropic) | | --model, -m | Model name | | --effort | Reasoning effort: low, medium, high or max (default: high) | | --print, -P | Run the prompt, print the result, exit | | --stream | Stream text to stdout in real time, with --print | | --output-schema | Require pipe-mode output to match an inline JSON Schema, or @file | | --permission | JSON tool permissions for run mode | | --max-turns | Cap agent/tool iterations in run mode — a safety limit for unattended CI work | | --resume, -r [id] | Resume a session; prefix match if an id is given | | --auto-accept, -y | Skip tool confirmations | | --deny-writes | Block all write operations | | --openai-api | OpenAI API mode: responses or chat (default: responses) | | --ollama-host / --ollama-port / --ollama-endpoint / --ollama-api-key | Ollama connection | | --help, -h / --version, -v | Help, version |

Slash commands

<details> <summary>29 commands, available in any session</summary>

| Command | What it does | | --- | --- | | /help | Show available commands | | /model | Show or change the current model | | /fast | Switch to a fast, lightweight model | | /deep | Switch to a powerful model for complex tasks | | /effort | Show or change reasoning effort | | /context | Show context window usage | | /compact | Compact conversation history to free up context | | /plan | Show, list, or update the active plan (/plan list, /plan <n> <status>, /plan clear) | | /steer | Add context or direction to guide the agent | | /undo | Revert the last file change | | /memory | Manage persistent memories | | /remember | Save a memory | | /forget | Delete a memory by name | | /resume | Resume a previous session | | /search | Search past sessions by keyword | | /branch | Fork a new session or list branches | | /name | Name the current session | | /export | Export conversation as a markdown file | | /new | Start a new chat without deleting saved sessions | | /clear | Start a new chat (alias: /new) | | /watch | Watch files and run a command on change | | /loop | Repeat a prompt on an interval | | /schedule | Manage persistent scheduled tasks | | /changelog | Show release notes for the current version | | /skills | Manage skills: list, install, remove, or browse the marketplace | | /agents | Manage service agents (list, install, create) | | /ps | Run a brief side query without interrupting the main task | | /debug | Show internal state for debugging | | /exit | Exit Agav |

</details>

Installation

Quick install (recommended)

One command, no Node.js required — Agav ships as a self-contained binary compiled with Bun:

curl -fsSL https://agav.dev/install.sh | bash

For Windows PowerShell:

irm https://www.agav.dev/install.ps1 | iex

[!NOTE] The www. is deliberate. agav.dev redirects with a 308, and Windows PowerShel