unleash

unleash your agent.

unleash is...

an agent CLI version manager. nvm for AI agents — claude code, codex, clanker code, antigravity, gemini, opencode, pi, hermes — with a rich TUI
a compatibility layer that lets you start in claude code, then continue where you left off in codex.
a unified cli that brings all your code agents under the same signature. No more confusion about claude -p vs codex run
an enabler for more advanced features such as self-restart, auto-mode and more.
Made for the sandbox — Use ours, Bring your own, or take the risk

Install

curl -fsSL unleash.software/install | bash

or with docker:

docker run --rm -it marksverdhei/unleash

(with auth tokens)

docker run -it --rm -e ANTHROPIC_API_KEY -e CLAUDE_CODE_OAUTH_TOKEN -e OPENAI_API_KEY -e GEMINI_API_KEY -e OPENROUTER_API_KEY marksverdhei/unleash

See Installation for build-from-source, platform details, and non-interactive mode.

unleash is best run in a sandbox. Bring your own or use ours — see the Docker + gVisor sandbox guide for more hardened containers with LAN isolation.

After install:

unleash          # Launch TUI (profiles, versions, settings)
unleash claude   # Start Claude with unleash features
unleash codex    # Start Codex with unleash features
unleash clanker  # Start Clanker Code with unleash features
unleash agy      # Start Antigravity CLI with unleash features
unleash gemini   # Start Gemini CLI with unleash features
unleash opencode # Start OpenCode with unleash features
unleash pi       # Start Pi with unleash features
unleash hermes   # Start Hermes Agent with unleash features

Run the same install command again to update to the latest version.

Custom Agent CLIs

Define any CLI agent in ~/.config/unleash/config.toml:

[[custom_agents]]
name = "aider"
binary = "aider"
description = "AI pair programming in your terminal"
github_repo = "paul-gauthier/aider"

[custom_agents.polyfill]
headless = { flag = "--message" }
session = { continue_strategy = { flag = "--restore-chat-history" }, resume_strategy = { flag = "--restore-chat-history" } }
fork = "unsupported"
model_flag = "--model"                    # required
yolo_flag = "--yes"                       # optional

Add a matching profile at ~/.config/unleash/profiles/aider.toml (any built-in profile makes a fine starting template), then launch with unleash aider — full polyfill support and TUI integration. See Custom Agent CLIs for the full schema, polyfill strategy reference, and multi-agent examples.

CLI Usage

Running Agents

unleash <profile> [unified flags] [-- agent-specific flags]

The first argument is always a profile name. The eight default profiles (claude, codex, clanker, agy, gemini, opencode, pi, hermes) map to their respective agents. Custom profiles can target any agent with custom settings — see docs/custom-agents.md.

unleash claude -m opus -c              # Continue last Claude session with Opus
unleash codex --safe                   # Run Codex with approval prompts
unleash clanker --name Cleo            # Run Clanker as the Chloe character alias
unleash gemini -p "fix the tests"     # Gemini headless mode
unleash hermes -m claude-sonnet-4      # Hermes with a model override
unleash work                           # Run a custom "work" profile

Unified Flags

These flags work identically across all agents. unleash translates them into the correct native syntax.

Flag	Short	Description	Default
`--safe`		Restore approval prompts (permissions bypassed by default)	off
`--prompt <prompt>`	`-p`	Run non-interactively with the given prompt
`--model <model>`	`-m`	Model to use for the session
`--continue`	`-c`	Continue the most recent session
`--resume [id]`	`-r`	Resume a session by ID, or open picker
`--fork`		Fork the session (use with `--continue` or `--resume`)
`--auto`	`-a`	Enable auto-mode (autonomous operation)

Anything after -- is passed directly to the agent CLI unchanged:

unleash claude -m opus -- --effort max --verbose
#      ^^^^^^ ^^^^^^^^    ^^^^^^^^^^^^^^^^^^^^^^^^^
#      Profile  Unified    Passthrough (Claude-specific)

How Translation Works

unleash	Claude	Codex	Clanker	Antigravity (`agy`)	Gemini	OpenCode	Pi	Hermes
`-p <prompt>`	`-p <prompt>`	`exec <prompt>`	`exec <prompt>`	`-p <prompt>`	`-p <prompt>`	`run <prompt>`	`-p <prompt>`	`-z <prompt>`
`-c`	`--continue`	`resume --last`	`resume --last`	`--continue`	`--resume latest`	`--continue`	`--continue`	`--continue`
`-r [id]`	`--resume [id]`	`resume [id]`	`resume [id]`	`--conversation [id]`	`--resume [id]`	`-s <id>`	`--session <id>`	`--resume [id]`
`--fork`	`--fork-session`	`fork` subcommand	`fork` subcommand	(unsupported)	(unsupported)	`--fork`	`--fork`	`--worktree`
(default)	`--dangerously-skip-permissions`	`--dangerously-bypass-approvals-and-sandbox`	`--dangerously-bypass-approvals-and-sandbox`	`--dangerously-skip-permissions`	`--yolo`	(no-op)	(no-op)	`--yolo`

Management Commands

unleash                    # Launch TUI
unleash update             # Update all agents (parallel progress bars)
unleash update --check     # Check for updates without installing
unleash update codex       # Update a specific agent
unleash update clanker     # Build and atomically install the current Clanker fork
unleash version            # Show installed versions
unleash version --list     # List available versions
unleash auth               # Check authentication status
unleash agents status      # Show all agent versions and update status
unleash stream --harness claude-code  # Normalize live harness JSONL from stdin
unleash serve claude --name work      # Serve one headful instance through OpenAI `/v1`

unleash stream exposes the same canonical live-event interface to shell and non-Rust consumers. It currently adapts Claude Code stream-json and Codex exec --json, preserves unknown frames, and emits a session_id join key on every output event. See the live stream normalization guide for the schema, attended/headless behavior, and upstream limitations.

unleash serve wraps one stateful agent behind /v1/models and /v1/chat/completions. Its primary mode keeps the harness headful and projects the same native session history to JSON/SSE; --headless is a resumed per-turn fallback. Live model IDs combine instance name, model slug, and harness. See the OpenAI-compatible gateway guide.

Version Management

unleash manages versions for all eight built-in agent CLIs:

Claude Code: Native binary (GCS) or npm install
Codex: Prebuilt binary from GitHub releases, cargo build fallback
Clanker Code: Fork-owned clanker branch source build with atomic install and revision receipt
Antigravity CLI (agy): AUR helper (yay/paru) on Arch; download from antigravity.google elsewhere
Gemini CLI: npm install
OpenCode: Built-in opencode upgrade command
Pi: npm install (@earendil-works/pi-coding-agent; updating migrates off the deprecated @mariozechner/pi-coding-agent)
Hermes Agent: Official curl install script (always installs latest)

Version filtering:

Blacklist mode (default for Claude): All versions allowed except known-bad ones
Whitelist mode (default for Codex): Only verified versions allowed
Version lists are maintained in Cargo.toml and compiled into the binary

Clanker updates are source builds from the current head of the private heiervang-technologies/clanker-code clanker branch. They require authenticated Git access plus Rust/Cargo, and currently have no signed release artifact or historical version picker. Unleash records the exact installed commit and treats that receipt as valid only while its owned binary still reports the recorded product version.

Extended Capabilities

Features that unleash adds on top of the base agent CLIs:

Available Now

Self-restart: Restart the agent while preserving session state (unleash-refresh)
Auto-mode: Autonomous operation via Stop hook + flag file system
Plugin system: Custom functionality loaded via --plugin-dir
MCP refresh: Detect MCP configuration changes and trigger reload
Voice output: Multi-provider TTS for agent responses (VibeVoice, OpenAI, ElevenLabs)
Profile system: Named configurations with per-agent settings, env vars, and themes
Parallel updates: Update all agents simultaneously with progress visualization
Skill synchronization: Sync custom agent skills and instruction sets across all harnesses (unleash skills sync)

Cross-CLI Session Crossload

Load conversation history from any agent into any other. Browse all sessions with unleash sessions, then crossload with -x:

unleash claude -x codex:rust-eng     # Load Codex session into Claude
unleash gemini -x claude:rice-chief  # Load Claude session into Gemini
unleash claude -x                    # Interactive session picker

Source → Target	Status
Claude → Gemini	🟢 Lossless
Gemini → Claude	🟢 Lossless
Codex → Claude	🟢 Lossless
Claude → Codex	🟢 Lossless
OpenCode → Claude	🟢 Lossless
Claude → OpenCode	🟡 Partial
Codex → Gemini	🟢 Lossless
Gemini → Codex	🟢 Lossless
OpenCode → Gemini	🟡 Partial
OpenCode → Codex	🟡 Partial
Gemini → OpenCode	🟡 Partial
Codex → OpenCode	🟡 Partial

🟢 Lossless · 🟡 Partial · ⚪ Pending — Full matrix

Pi / Hermes / Antigravity (agy) are wired but not in the headline matrix yet — see Additional CLIs.

Cross-CLI Skill Synchronization

Sync custom agent skills, custom commands, and references across all your harnesses. Define a skill once in Claude or OpenCode, and run everywhere:

unleash skills sync                 # Sync from source to all targets
unleash skills list                 # List active skills across harnesses
unleash skills status               # Show the status matrix

See Skill Synchronization and the Fidelity Matrix for full details.

On the Roadmap

Directory navigation and workspace management
PTY terminal middleware for session scripting

Profiles

Profiles are TOML files in ~/.config/unleash/profiles/. Each profile specifies an agent CLI, arguments, environment variables, and theme.

# ~/.config/unleash/profiles/work.toml
name = "work"
description = "Work profile with Claude"
agent_cli_path = "claude"
agent_args = []
theme = "blue"

[env]
ANTHROPIC_API_KEY = "sk-..."

Per-agent overrides allow a single profile to customize behavior for different agents:

[agents.claude]
extra_args = ["--effort", "high"]

[agents.codex]
extra_args = ["--full-auto"]

TUI

The TUI (unleash with no arguments) provides:

Profile management: Create, edit, duplicate, search profiles
Version management: Browse, install, and switch agent versions
Settings: Auto-update toggles, theme selection, animations

Navigate with j/k or arrows, Enter to select, Esc to go back, ? for help.

Plugins

All extended functionality is implemented as plugins in plugins/bundled/:

Plugin	Description
auto-mode	Autonomous operation via Stop hook enforcement
mcp-refresh	Detect MCP config changes and notify for reload
process-restart	Self-restart with session preservation
hyprland-focus	Window transparency on Hyprland during agent work
omnihook	Unified hook handler with voice input integration
skillsync	Cross-CLI skill & prompt synchronization

Creating Plugins

mkdir -p plugins/my-plugin

// plugins/my-plugin/plugin.json
{
  "name": "my-plugin",
  "version": "1.0.0",
  "description": "What it does",
  "hooks": {
    "Stop": "./hooks/stop.sh"
  }
}

See the plugin development guide for details.

Installation

One-liner (recommended)

curl -fsSL unleash.software/install | bash

Downloads a prebuilt binary for your platform (Linux x86_64/aarch64, macOS x86_64/aarch64), installs to ~/.local/bin, and launches the interactive setup to pick your default agent.

For non-interactive installs (CI, scripts):

curl -fsSL unleash.software/install | bash -s -- --boring

Docker (Sandboxed)

# One-time setup (installs gVisor, network isolation, pulls image)
sudo unleash sandbox setup

# Run an agent
unleash sandbox run claude

Or run the image directly:

docker run -it --rm -e ANTHROPIC_API_KEY marksverdhei/unleash

All 5 agent CLIs are pre-installed. See the Docker + gVisor sandbox guide for hardened setups with LAN isolation and named sandboxes.

CUDA / GPU Cloud

For GPU workloads, use the CUDA variant with CUDA 12.8, PyTorch, and all agent CLIs:

docker run --gpus all -it --rm marksverdhei/unleash:cuda

See the Docker guide for Vast.ai deployment and compose overrides.

Build from source

Requires Rust:

git clone https://github.com/heiervang-technologies/unleash.git
cd unleash
cargo build --release
./scripts/install.sh

Or force the installer to build from source instead of downloading:

BUILD_FROM_SOURCE=1 bash <(curl -fsSL unleash.software/install)

Headless build (no TUI)

The interactive splash / agent picker (the splash binary) links ratatui and crossterm behind the default tui feature. In headless environments you can build the unleash CLI without them:

cargo build --no-default-features

This drops the tui feature and the splash binary; the unleash CLI and all of its subcommands (version management, crossload, search, sandbox, …) still build and run. CI enforces this contract via cargo check --no-default-features so it cannot silently regress.

Platform support

Platform	Binary	Method
Linux x86_64	`unleash-linux-x86_64`	Static musl binary
Linux aarch64	`unleash-linux-aarch64`	Static musl binary
macOS x86_64	`unleash-macos-x86_64`	Native binary
macOS aarch64	`unleash-macos-aarch64`	Native binary

Linux binaries are statically linked (musl) — no glibc dependency. Works on any Linux including WSL, Alpine, and containers.

Agent CLI dependencies

unleash itself has no dependencies beyond curl. Agent CLIs have their own:

Agent	Install method	Requires
Claude Code	Native binary (GCS)	curl
Codex	Prebuilt binary (GitHub)	curl
Gemini CLI	npm	Node.js
OpenCode	Built-in upgrade / npm fallback	curl (or Node.js)

If npm is missing when installing Gemini or OpenCode, unleash will offer to install Node.js via nvm.

Authentication

Each agent CLI uses its own API key:

# Set keys as environment variables
export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...
export GEMINI_API_KEY=...

# Or use OAuth (Claude Code)
claude login

Verify with unleash auth or unleash auth --verbose.

Architecture

unleash/
├── src/                    # Rust CLI & TUI
│   ├── cli.rs             # Argument parsing + polyfill flags
│   ├── polyfill.rs        # Unified flag → agent-specific translation
│   ├── launcher.rs        # Agent wrapper with restart/auto-mode
│   ├── updater.rs         # Parallel update orchestrator
│   ├── progress.rs        # Terminal progress bar renderer
│   ├── agents.rs          # Agent definitions + version management
│   ├── config.rs          # Profile + settings management
│   └── tui/               # Terminal UI (ratatui)
├── plugins/bundled/        # Plugin extensions
├── scripts/                # Install/uninstall scripts
└── docs/                   # Specs and guides

Contributing

Contributions are very welcome! See CONTRIBUTING.md for guidelines.

Documentation

Links

Issue Tracker
Claude Code | Codex | Gemini CLI | OpenCode

Built by Heiervang Technologies

Name		Name	Last commit message	Last commit date
.github		.github
assets		assets
data		data
docker		docker
docs		docs
examples		examples
plugins/bundled		plugins/bundled
scripts		scripts
src		src
tests		tests
.dockerignore		.dockerignore
.gitignore		.gitignore
.release-please-manifest.json		.release-please-manifest.json
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CONTRIBUTING.md		CONTRIBUTING.md
Cargo.lock		Cargo.lock
Cargo.toml		Cargo.toml
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
build.rs		build.rs
deny.toml		deny.toml
release-please-config.json		release-please-config.json

Navigation Menu

Uh oh!

Repository files navigation

unleash

Install

Custom Agent CLIs

CLI Usage

Running Agents

Unified Flags

How Translation Works

Management Commands

Version Management

Extended Capabilities

Available Now

Cross-CLI Session Crossload

Cross-CLI Skill Synchronization

On the Roadmap

Profiles

TUI

Plugins

Creating Plugins

Installation

One-liner (recommended)

Docker (Sandboxed)

CUDA / GPU Cloud

Build from source

Headless build (no TUI)

Platform support

Agent CLI dependencies

Authentication

Architecture

Contributing

Documentation

Links

About

Releases

Packages

Contributors

Languages

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

unleash

Install

Custom Agent CLIs

CLI Usage

Running Agents

Unified Flags

How Translation Works

Management Commands

Version Management

Extended Capabilities

Available Now

Cross-CLI Session Crossload

Cross-CLI Skill Synchronization

On the Roadmap

Profiles

TUI

Plugins

Creating Plugins

Installation

One-liner (recommended)

Docker (Sandboxed)

CUDA / GPU Cloud

Build from source

Headless build (no TUI)

Platform support

Agent CLI dependencies

Authentication

Architecture

Contributing

Documentation

Links

About

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages