vtcode

Semantic AI coding agent with multi-provider LLM support

TLDR

Launch the interactive coding agent in the current directory

$ vtcode

Ask the agent a one-off question and exit

$ vtcode ask "[explain the project structure]"

Execute a shell command through the agent's safe runner

$ vtcode exec "[cargo test]"

Use a specific model provider and model

$ vtcode --provider [anthropic] --model [claude-sonnet-4-5]

Run against an alternate workspace directory

$ vtcode --workspace [path/to/repo]

Open the configuration file for editing

$ vtcode config edit

Print the resolved configuration

$ vtcode config show

Authenticate with a provider via OAuth

$ vtcode /login

SYNOPSIS

vtcode [global-flags] [subcommand] [args]

vtcode is an open-source semantic coding agent written in Rust. It connects to one or more LLM providers and exposes a tool-using assistant that can read, edit, and execute code inside a workspace. A Tree-sitter-based parser gives the agent structural understanding of source files in many languages, enabling targeted edits rather than blind text rewrites.Tool execution is gated by a configurable policy: each tool can be set to allow, deny, or prompt, and shell commands are statically validated with tree-sitter-bash before being run. Sandboxing keeps file operations within the configured workspace boundary.Multi-provider support — OpenAI, Anthropic, DeepSeek, Gemini, Z.AI, Moonshot, OpenRouter, MiniMax, Ollama, LM Studio — is wired through a single configuration file with automatic failover, prompt caching, and context-window-aware truncation.

PARAMETERS

ask prompt

Send a single prompt to the configured model and print the response.

exec command

Run command through the agent's policy-aware shell runner with tree-sitter-bash validation.

config [show|edit|path]

Inspect or edit the vtcode.toml configuration.

--provider name

Override the default LLM provider (e.g. openai, anthropic, deepseek, gemini, ollama, lmstudio).

--model id

Override the default model id for the selected provider.

--workspace path

Set the working directory the agent operates in (defaults to $PWD).

--config file

Use file instead of the default vtcode.toml.

--no-tools

Disable tool execution; the agent only emits text.

--yes

Auto-approve tool calls instead of prompting per action.

-v, --verbose

Increase log verbosity.

--version

Print the version and exit.

-h, --help

Show usage information.

CONFIGURATION

User configuration lives in vtcode.toml (project-level) or ~/.config/vtcode/vtcode.toml (global). Common keys:

[providers.<name>] — API base URL and credential source for each provider.

[oauth] — OAuth client settings; tokens are stored in the OS keychain.

[tools] — per-tool allow/deny/prompt policies.

[workspace] — root path and ignore globs.

[security] — shell sandbox, command allow/deny lists, max output size.

[context] — context window budget and truncation strategy.

[hooks] — shell commands to run on lifecycle events (session start, tool call, etc.).

Secrets are typically read from environment variables (e.g. OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY) or from the OS-native credential store.

SLASH COMMANDS

/login

Run an OAuth flow to authenticate with the active provider.

/agent

Inspect delegated subagents in the current session.

/agents active

List currently running subagents.

/subprocesses

Open the local agents drawer (also bound to Alt+S).

KEYBOARD SHORTCUTS

$ Ctrl+B   Toggle background subagent
Alt+S    Open Local Agents drawer
Ctrl+C   Cancel current request

CAVEATS

vtcode is pre-1.0 — flags, configuration keys, and slash commands change between releases. LLM providers may charge per token; review [security] policies before enabling auto-approve (--yes) on a real codebase. Some features (Zed editor integration, A2A protocol, sandbox isolation) require additional setup beyond the default install.

HISTORY

vtcode is developed by Vinh Nguyen (vinhnx) and published at github.com/vinhnx/vtcode. It is distributed via Cargo (cargo install vtcode), Homebrew (brew install vtcode), and an install script. Releases are published under an open-source license on GitHub.