The Problem with Every Other AI Coding Tool
Every popular AI coding assistant shares the same architectural assumption: you are sitting in front of the machine running your code. The tool runs locally, the context is local, and the session lives and dies with that terminal window.
That assumption breaks the moment reality intervenes. You close your laptop mid-session. You connect from a different machine. Your team lead wants to watch a debugging session in real time. A security policy prevents your source code from touching a third-party cloud endpoint. Your production environment is air-gapped.
Most tools respond to these situations with silence. The session is gone. Start over.
There is a deeper problem too. Every AI coding assistant that runs as a browser extension, a cloud IDE, or a SaaS subscription sends your code somewhere. The inference happens on infrastructure you do not control, under privacy policies written by someone else’s legal team. For regulated industries, defense contractors, financial institutions, and anyone handling client intellectual property, it is not acceptable.
Anvil v2.2.4 was built around a different set of assumptions.
Who Anvil Is Built For
Anvil is a terminal-native AI coding agent written in Rust. It is designed for three overlapping audiences.
Security-conscious teams who cannot or will not send source code through a third-party cloud pipeline. Anvil supports fully local inference via Ollama. When you run a local model, your code never leaves the machine. No telemetry. No retention. No exceptions.
DevOps and platform engineers who live in the terminal and expect their tools to behave like terminal tools — keyboard-driven, composable, scriptable, and fast. Not a browser tab masquerading as a development environment.
Distributed teams that need to pair-program with AI across devices and time zones. The remote control feature lets any authorized device observe and interact with a live Anvil session over a secure WebSocket relay — without exposing your machine to the internet.
What Makes v2.2.4 Different
Live Remote Control
This is the feature no other terminal AI assistant offers. Type /remote-control and Anvil generates a unique session URL backed by a WebSocket relay. Any browser — on any device, anywhere — can connect to that URL and see your live session in real time: the conversation history, tool calls in progress, model responses as they stream.
The relay uses a pairing code for initial authentication. Session hashes are randomly generated per-connection. Remote control enables real-time code review, async pair programming, and incident response scenarios where a senior engineer needs to watch a junior engineer’s AI session without being on the same machine.
Typed Credential Vault
Anvil ships with an AES-256-GCM encrypted credential vault using Argon2id key derivation. The vault supports 21 typed credential categories: API keys, SSH keypairs, database credentials, TLS certificates, TOTP secrets, cloud credentials, Docker registries, Kubernetes configs, and more.
TOTP codes are generated live from stored secrets. Credential scanning (/vault scan) sweeps environment variables, dotfiles, and .env files and offers to vault anything that looks like a secret. The master password never touches disk.
37 Status Line Widgets with Visual Editor
The Anvil status bar is fully configurable through a visual editor. Choose from 37 available widgets — model info, token economics, burn rate, context window, git status, remote control indicators, MCP server count — and arrange them across multiple status lines. 16 presets included, from minimal zen to maximalist dashboard.
5 AI Providers with Configurable Failover
Anvil supports Anthropic (Claude), OpenAI (GPT-4o, o3), Google Gemini, xAI (Grok), and Ollama (any local model). Configure a primary provider and a failover chain with per-model cooldown timers and cost budgets. When the primary model hits a rate limit, Anvil promotes the next model automatically.
Security Hardening
v2.2.4 completes a security hardening pass across the entire codebase: constant-time HMAC verification, plugin command injection prevention, path traversal protection, cryptographic session IDs, network egress control, credential redaction in logs, and signed audit transcripts.
Zero-Warning Codebase, 428 Tests
The entire workspace compiles with zero warnings under clippy::pedantic. 428 tests across 9 crates with zero failures and zero flakes. 110 functions made const fn. Redundant clones eliminated. main.rs modularized into focused modules.
Honest Comparison
Anvil vs. Claude Code
Anvil advantages: Live remote control, typed credential vault with TOTP, configurable status bar, multi-provider failover, Ollama local inference, network egress control, signed audit transcripts.
Claude Code advantages: Deeper Anthropic model integration, git worktree isolation for agents, mature MCP ecosystem with hundreds of community servers. If your workflow is Anthropic-only and you want the richest MCP plugin surface, Claude Code is the better-integrated choice.
Anvil vs. OpenAI Codex
Codex is an API product, not a terminal environment. It processes requests through OpenAI’s cloud infrastructure with no offline option. Anvil runs entirely on your machine with local inference support.
Anvil vs. Cursor / Continue
Cursor and Continue are IDE extensions — the right choice if your workflow is IDE-centered. Anvil is a standalone terminal environment that runs over SSH, in tmux, and in contexts where a GUI IDE is unavailable.
Getting Started
brew install culpur/anvil/anvilOr download directly from github.com/culpur/anvil/releases/latest for macOS, Linux, and Windows.
anvil # Start interactive session
/remote-control # Share via browser
/vault init # Initialize credential vault
/configure statusline # Customize status bar
/productivity # Session statsWhat’s Next
Session persistence and reconnect. Sessions that survive TUI restarts and allow async hand-off between team members.
Deeper plugin system. Plugins that register slash commands, inject context, and interact with the credential vault.
Structured agent output. JSON output mode for CI/CD integration.
Anvil v2.2.4 is available now at culpur.net/anvil.
