Beyond

An AI-native terminal written in Rust. Beyond replaces the traditional scroll buffer with a block-oriented model (Warp-like) — every piece of content (shell output, agent messages, approvals, diffs, tool calls) is a persistent, addressable Block with provenance. Agents are first-class, long-lived processes with capability sets and resource limits, supervised like OS processes. Rendering is GPU-accelerated (wgpu + glyphon) inside a single winit window.

Status: early. Phase 2 (glyphon 0.8 text, wgpu 24, continuous render loop) has landed. Expect breaking changes.

Highlights

• Blocks, not scrollback. Shell commands, agent turns, and approvals are all Blocks with ULID IDs, parent/child relationships, and a provenance chain.
• Agents as processes. The AgentSupervisor spawns, monitors, and kills agents. Each has a CapabilitySet and ResourceLimits; tool calls flow through a CapabilityBroker.
• Multi-provider LLM support. Ollama (local + Cloud/Turbo), llama.cpp, and Apple MLX — switch at runtime with /provider.
• GPU rendering. Single winit window, wgpu 24, glyphon 0.8 for text, custom per-block renderers.
• OSC-133 shell integration. zsh/bash hooks let Beyond see where one command ends and the next begins.

Requirements

• A GPU that supports wgpu (Metal on macOS, Vulkan/DX12 elsewhere)
• At least one LLM provider running for agent features:
  • Ollama — ollama serve (local) or Ollama Cloud (Turbo)
  • llama.cpp — llama-server -m model.gguf --jinja -c 8192
  • Apple MLX — mlx_lm.server --model <id> (macOS Apple Silicon)

Installation

Quick install (macOS / Linux)

curl -fsSL https://raw.githubusercontent.com/PewterZz/Beyond/main/install.sh | bash

Or specify a version:

VERSION=0.1.0 curl -fsSL https://raw.githubusercontent.com/PewterZz/Beyond/main/install.sh | bash

Cargo (from source)

cargo install beyondtty

Homebrew (macOS / Linux)

brew install PewterZz/tap/beyondtty

Build from source

git clone git@github.com:PewterZz/Beyond.git
cd Beyond
cargo build --release
# Binaries at target/release/beyondtty and target/release/beyonder

Quick start

beyondtty   # or beyonder — both work

The window opens at 1280x800. Type shell commands normally, or /help for slash commands. /agent spawn <name> to start an agent.

Common commands

cargo build                      # debug build
cargo build --release            # release (LTO, codegen-units=1)
cargo run                        # launch the app
cargo check -p <crate>           # fast single-crate check
cargo test                       # all tests
cargo test -p beyonder-core      # single crate
cargo clippy --workspace --all-targets
cargo fmt --all

Logging uses tracing. Tune with RUST_LOG (default beyonder=info,wgpu_core=warn,wgpu_hal=warn). Logs go to stderr; redirect with cargo run 2> run.log.

Architecture

The workspace binary is a thin winit::ApplicationHandler that owns the tokio runtime and delegates to beyonder-ui::App. Crates form a layered graph:

Crate	Role
beyonder-core	Pure data model: Block, AgentInfo, SessionId, CapabilitySet, ProvenanceChain. No I/O.
beyonder-store	SQLite persistence (bundled rusqlite): BlockStore, SessionStore, migrations.
beyonder-terminal	PTY (portable-pty) + terminal emulation (alacritty_terminal), OSC-133 shell hooks, BlockBuilder.
beyonder-acp	Agent Client Protocol: messages, transport, AcpClient.
beyonder-runtime	AgentSupervisor, CapabilityBroker, tool registry, AgentBackend trait + Ollama/llama.cpp/MLX providers.
beyonder-gpu	wgpu renderer, glyphon text atlas, per-block renderers, viewport/scrolling.
beyonder-ui	App wiring, input editor, slash commands, history, mode detection.
beyonder-config	TOML config loader (BeyonderConfig).

State-advancement (App::tick()) runs in winit's about_to_wait, not inside RedrawRequested — macOS suppresses redraws for hidden windows, and we don't want streaming agent output to stall when the window is occluded.

See CLAUDE.md for a deeper tour written for AI coding assistants working in this repo.

Configuration

Config lives at $XDG_CONFIG_HOME/beyond/config.toml (or the platform equivalent). Missing or malformed config falls back to defaults — see BeyonderConfig::load_or_default.

LLM Providers

# Ollama (default)
[provider]
kind = "ollama"
base_url = "http://localhost:11434"       # optional; cloud: "https://ollama.com"
api_key_env = "OLLAMA_API_KEY"            # optional; omit for local

# llama.cpp
[provider]
kind = "llama_cpp"
base_url = "http://127.0.0.1:8080/v1"

# Apple MLX (macOS Apple Silicon)
[provider]
kind = "mlx"
base_url = "http://127.0.0.1:8080/v1"

Switch at runtime with /provider ollama|llama_cpp|mlx and /model <name>. Changes take effect on the next agent spawn.

Contributing

See CONTRIBUTING.md and CODE_OF_CONDUCT.md. In short: keep PRs focused, run cargo fmt and cargo clippy --workspace --all-targets before pushing, and add tests for behavior changes.

License

GPL-3.0-or-later. See LICENSE.