>anvi.

Terminal-native AI coding agent. Built in Rust. Powered by Gemini.

Anvi is a terminal AI coding agent in the same family as Claude Code and Cursor's agent workflows, but designed as a native Rust CLI instead of a web shell wrapped in a terminal. It gives you multi-turn tool use, code editing, shell execution, sub-agent orchestration, persistent memory, and git-aware workflows with fast startup, Gemini-native defaults, and an extension model built around installable skills, MCP servers, and plugins.

Anvi is built for people who want an agent that feels like part of the terminal: fast to launch, easy to script, comfortable in large repos, and extensible without rebuilding the core product.

• Startup target: sub-150ms cold start
• Model-first design: Gemini 3.x as the default provider
• Extension surface: skills, MCP servers, plugins, templates, memory
• Workspace model: local-first, project-aware, and git-aware

See Getting Started for a guided walkthrough, or Roadmap for the longer product vision.

Install

# npm (recommended)
npm install -g @anvi-cli/anvi

# Shell installer
curl -fsSL https://raw.githubusercontent.com/eagleisbatman/anvi/main/install.sh | sh

# Build from source
cargo install --git https://github.com/eagleisbatman/anvi.git --package anvi-tui --bin anvi --force

Quick Start

# Set API key
export GEMINI_API_KEY="your-key-here"

# Run
anvi

If you prefer another provider, Anvi also supports OPENAI_API_KEY and ANTHROPIC_API_KEY.

Features

• Multi-turn agent loop with 23+ built-in tools for files, shell, git, web, memory, and tasks
• /commit, /pr, /test-fix for AI-assisted git workflows and fix-until-green loops
• /plan -x for generating a plan and executing it step by step
• /review, /security, /architecture, and 14 more built-in specialist agents
• /skills search|install|create for browsing community skills or authoring your own
• /mcp browse|install|create for adding Model Context Protocol servers
• /suggest for project stack detection with skills and MCP recommendations
• /template for reusable prompt templates with variable expansion
• /memory for persistent memory, #keyword recall, and team-sharing flows
• /marketplace search|install|info for unified browsing and installing skills, MCP servers, and plugins
• Plugin system for custom tools in any language via ~/.anvi/plugins
• RTK integration for 60-90% token reduction on supported shell output
• CLI-Anything integration for controlling professional desktop software from the terminal
• @file mentions for prompt-time file injection and #keyword memory lookup
• Progress indicators, cost tracking, context budget visibility, and session summaries

Configuration

Anvi stores global config in ~/.anvi/config.toml and can layer project-local helpers under .anvi/.

# ~/.anvi/config.toml
[model]
provider = "gemini" # gemini | openai | anthropic
model_id = "gemini-3-flash-preview"

[optimization]
rtk_enabled = false # Install RTK and flip to true for token savings

Common provider environment variables:

• Gemini: GEMINI_API_KEY
• OpenAI: OPENAI_API_KEY
• Anthropic: ANTHROPIC_API_KEY

Useful companion settings:

• [permissions] mode = "default" for approval behavior
• theme = "dark" for terminal theme preference
• [[model.fallback]] for provider/model fallback chains

Commands Quick Reference

Slash commands are the main control surface inside the REPL. Specialist agents are also exposed as commands, so /review, /security, /scaffold, /debug, and friends behave like first-class built-ins.

| Command | Description | | --- | --- | | /new | Start a fresh session in the current workspace | | /sessions | Browse and resume recent sessions interactively | | /resume <id> | Resume a specific session by thread ID | | /compact | Summarize older context to free tokens | | /rewind | Remove the last user/assistant exchange | | /model [name] | Show or switch the active model | | /models | Browse built-in model options | | /cost | Estimate current session cost from token usage | | /read <path> | Read a file into the terminal | | /write <path> [content] | Write content or open a file in $EDITOR | | /edit <path> | Open a file in your configured editor | | /bash <cmd> | Run a shell command through the agent runtime | | /glob <pattern> | Find files by glob pattern | | /grep <pattern> | Search repository content with regex | | /plan [-x] <task> | Generate a plan, optionally auto-execute it | | /diff [--staged] | Render the current git diff | | /memory ... | List, search, export, import, share, or prune memory | | /skills ... | List, search, install, remove, or create skills | | /marketplace ... | Search, install, and inspect skills, MCP servers, and plugins | | /template ... | Run a reusable prompt template | | /commit | Generate a conventional commit message and commit | | /pr [base] | Draft and create a GitHub pull request | | /test-fix [n|--cmd ...] | Run tests, fix failures, and retry | | /undo | Revert uncommitted changes after confirmation | | /status | Show session, environment, memory, and runtime status | | /context | Inspect context budget and compaction pressure | | /tokens | Show input, output, and total token usage | | /init | Create project-local .anvi helper files | | /suggest | Detect the project stack and recommend add-ons | | /plugins | List discovered external plugin tools | | /theme [name] | Show or switch the terminal theme | | /login | Configure the active provider API key | | /logout | Clear the active provider API key | | /mcp ... | Browse, install, add, create, login, or remove MCP servers | | /permissions [mode] | Change approval mode for tools and edits | | /bg <prompt> | Run a prompt as a background job | | /fg [job-id] | Show output from a finished background job | | /jobs | List background jobs | | /help | Show the full command list |

Prompt shortcuts:

• @path/to/file.rs injects a file into the next prompt
• @path/to/dir/ injects a directory listing
• #keyword recalls matching memory entries into context

Skills & Plugins

Anvi has two extension layers:

• Skills are prompt-level behaviors. Built-in examples include /review, /security, /scaffold, /docs, /debug, and /onboard. Community skills can be searched and installed with /skills search <query> and /skills install <name>.
• Plugins are executable tools discovered from ~/.anvi/plugins. Each plugin is a trusted local binary plus a JSON manifest, so you can add custom tools in Python, shell, Ruby, or anything else that can read JSON on stdin and print JSON on stdout.

Examples:

/skills search ci
/skills install ci-pipeline
/skills create release-notes

/mcp browse
/mcp install github
/mcp create jira-sync

For hands-on setup, see docs/GETTING_STARTED.md.

Architecture

Anvi is a 6-crate Rust workspace:

anvi
├── mastra-core       shared agent/runtime types, streaming, workflows
├── mastra-memory     SQLite-backed sessions, tasks, and memory
├── mastra-providers  Gemini, OpenAI, and Anthropic provider adapters
├── mastra-tools      built-in tools, plugins, MCP, RTK, CLI-Anything
├── anvi-agent        orchestration, planning, config, skills, approvals
└── anvi-tui          CLI binary, REPL, commands, themes, rendering

Rough dependency flow:

anvi-tui
   -> anvi-agent
      -> mastra-core
      -> mastra-memory
      -> mastra-providers
      -> mastra-tools

Installation Notes

• Requires a Rust toolchain with cargo
• First run creates ~/.anvi/ automatically
• anvi --doctor checks API key, provider connectivity, memory, git context, and RTK status
• anvi -c resumes your most recent session
• anvi "prompt" runs one-shot mode without entering the full REPL

License

Apache 2.0