LINCE

Linux Intelligent Native Coding Environment — lince.sh

A toolkit that turns your terminal into a multi-agent engineering workstation — spawn parallel AI coding agents, track their status in real time, relay voice commands, all from a single TUI dashboard running in Zellij.

Demo

The full workflow: voice command → Whisper transcription → Claude Code executes in sandbox → backlog updates — all from one terminal.

The Dashboard

The primary way to use LINCE is the TUI Dashboard — a Zellij WASM plugin that acts as a command center for multiple Claude Code agents.

┌─────────────────────────────────────────────────────────┐
│  LINCE Dashboard                                        │
│ ┌───┬────────────┬──────────┬────────┬───────┐          │
│ │ # │ Name       │ Status   │Profile │Project│          │
│ ├───┼────────────┼──────────┼────────┼───────┤          │
│ │ 1 │ backend    │ Running  │ vertex │ api/  │          │
│ │>2 │ frontend   │ INPUT    │        │ web/  │          │
│ │ 3 │ tester     │   -      │ zai    │ tests/│          │
│ └───┴────────────┴──────────┴────────┴───────┘          │
├───────────────────────┬─────────────────────────────────┤
│                       │                                 │
│   VoxCode             │   Shell                         │
│   Voice input →       │                                 │
│   relayed to agents   │                                 │
│                       │                                 │
└───────────────────────┴─────────────────────────────────┘

What the dashboard gives you:

• Multi-agent: Spawn up to 8 AI coding agents in parallel (Claude Code, Codex, Gemini, OpenCode, Aider, and any custom agent), each in its own sandboxed pane
• Real-time status: See at a glance which agents are running, waiting for input, asking for permission, or stopped — five canonical states, color-coded
• Pane control: Show/hide agent panes with a keystroke (f to focus, h to hide)
• Voice relay: VoxCode transcriptions are piped directly to the focused agent
• Session persistence: Save/restore your agent constellation across sessions (Q to save & quit)
• Swimlane grouping: Agents auto-grouped by project directory when working across repos
• Sandbox isolation: Every agent runs inside agent-sandbox — full autonomy, zero host risk

Quick Start

For a detailed step-by-step guide with scenarios (Mini/Full/Custom), troubleshooting, and quick reference, see QUICKSTART.md.

Prerequisites

• Linux (tested on Fedora 43, works on Ubuntu/Debian/Arch)
• Zellij >= 0.40
• Claude Code (npm install -g @anthropic-ai/claude-code)
• bubblewrap (sudo dnf install bubblewrap / sudo apt install bubblewrap)
• Rust with wasm32-wasip1 target (for building the dashboard plugin)
• macOS (experimental): native Seatbelt sandbox backend (sandbox-exec, built into macOS — no extra install), Python 3.11+, Zellij. The legacy nono backend is deprecated — see the migration guide

Step 1: Install the sandbox

cd sandbox
./install.sh

This creates an isolated environment where Claude has full write access to your project but physically cannot reach your SSH keys, cloud credentials, or anything outside the project directory. See sandbox/README.md for the full configuration reference.

Step 2: Install the dashboard

cd lince-dashboard
./install.sh

The installer builds the WASM plugin, copies it to ~/.config/zellij/plugins/, installs Zellij layouts, sets up Claude Code status hooks, and creates the lince shell alias. After sourcing your shell config:

source ~/.bashrc
lince    # launch the dashboard

Step 3: Launch

lince

Press n to spawn an agent (quick), or N for the full wizard (name, sandbox profile, project directory). Press ? for the full keybindings overlay.

Alternative: Use the Quickstart Installer

For an interactive installer that handles all modules with dependency resolution:

./quickstart.sh

See QUICKSTART.md for more options (--mini, --full, --yes flags).

The Workflow in Practice

You (speaking):  "Pick up the next task from the backlog and start working on it"
                        │
                        ▼
VoxCode:         Whisper transcribes locally → pipes text to dashboard
                        │
                        ▼
Dashboard:       Routes text to focused/selected agent
                        │
                        ▼
Claude Code:     Reads the backlog via MCP → picks a task →
                 marks it in-progress → writes code → runs tests
                        │
                        ▼
Dashboard:       Status updates in real-time (Running → INPUT → Stopped)
                 Five canonical states: - / Running / INPUT / PERMISSION / Stopped

Modules

lince-dashboard/

The multi-agent TUI dashboard — a Zellij WASM plugin (Rust, ~900 KB) that manages multiple AI coding agents (Claude Code, Codex, Gemini, OpenCode, and any custom agent). Spawn agents, monitor status, show/hide panes, relay voice input, persist sessions. Agent types are fully config-driven — add new agents via TOML or use the /lince-add-supported-agent skill. See lince-dashboard/README.md.

Documentation: Usage Guide | Configuration | Agent Examples

lince-config/

Structured CLI for reading and editing LINCE configuration files (~/.agent-sandbox/config.toml and ~/.config/lince-dashboard/config.toml). Preserves comments and formatting via tomlkit. Installed to ~/.local/bin/lince-config.

Also powers the /lince-configure skill — a natural-language interface that lets any AI coding agent read and modify its own configuration interactively (conversational or guided-menu mode). Installed automatically by quickstart.sh or lince-dashboard/install.sh.

Config v2: configuration is converging on a single versioned policy file (~/.config/lince/lince.toml) + a shipped agent registry. Get started with lince-config discover / lince-config apply <agent>+<level>+<provider>; existing installs keep working unchanged — see the user migration guide and the developer migration guide.

sandbox/

Bubblewrap-based sandbox for running AI coding agents safely. Restricts filesystem access, blocks git push, isolates environment variables, and hides host processes — with near-zero overhead. Supports any agent via --agent flag (agent-sandbox run -a codex, -a gemini, etc.). Used by the dashboard to spawn every agent.

Documentation: CLI Reference | Configuration | Security Model

/lince-add-supported-agent skill (bundled with lince-dashboard)

An agentskills.io-compliant skill that lets any AI coding agent register itself with the lince ecosystem. Generates correct TOML configuration for both agent-sandbox and lince-dashboard. Installed automatically by lince-dashboard/install.sh.

/lince-configure skill (bundled with lince-dashboard)

An agentskills.io-compliant skill for natural-language configuration of LINCE. Ask any AI agent to configure providers, change sandbox levels, set API keys, adjust dashboard settings, or diagnose issues — it drives lince-config under the hood. Supports conversational and guided-menu interaction. Installed automatically by lince-dashboard/install.sh (requires the lince-config CLI).

Related Projects

Project	Description
VoxCode	Voice input for AI agents — local Whisper transcription, integrates with the dashboard via Zellij pipes
VoxTTS	Text-to-Speech with local GPU/CPU engines (Kokoro, Piper)
Agent Ready Skill	Assess any project's readiness for agentic coding (agentskills.io)

License

MIT