github.com/newjordan/angelX

Follow Your Star —
Discover the Unknown

A self referential world model harness built for customization and long form agentic enjoyment.

✦

ONE-COMMAND LINUX INSTALL · x86_64 · BUILDS FROM SOURCE ON FIRST RUN

git clone https://github.com/newjordan/angelX.git && cd angelX && ANGEL_VIDEO=0 ./bin/angelX
rust
ratatui
v8
world
node
python
git
docker
linux
ffmpeg
playwright
opencode
dotagents
models

AngelX created by Jordan Newman @frostforger

◇ measured · the race to 54 · 2026-09-21 ◇

DEEPSEEK V4.1 FLASH · THINKING OFF

GLM-5.3-FLASH · THINKING LOW

angelXOpenCodeomp
◇ measured · graded attempts · 2026-09-21 ◇
◇ measured · output per task · 2026-09-21 ◇
◇ measured · seconds per attempt · 2026-09-21 ◇

DEEPSEEK V4.1 FLASH · THINKING OFF

GLM-5.3-FLASH · THINKING LOW

■ angelX+ OpenCode× omp
◇ measured · context burned · 2026-09-21 ◇

DEEPSEEK V4.1 FLASH · THINKING OFF

GLM-5.3-FLASH · THINKING LOW

angelXOpenCodeomp
◇ measured · cache · 2026-09-21 ◇

DEEPSEEK V4.1 FLASH · THINKING OFF

GLM-5.3-FLASH · THINKING LOW

angelXOpenCodeomp

* All tests performed on the Prime Intellect evaluators (Verifiers v0.3.1) · prime-quality-v1: 18 repository-repair tasks (8 JS, 7 Python, 3 Rust), 3 runs per harness per model, 324 graded attempts, pass/fail decided by each task’s own tests · angelX a5e788f · oh-my-pi 18.2.4 · opencode 1.18.31 · DeepSeek V4.1 Flash, thinking off · GLM-5.3-Flash, thinking low (the model’s floor) · temperature 0 · 8,192-token output cap · fresh environment per attempt · 2026-09-21

A one of a kind TUI experience.

Multi-panel TUI with unique and customizable model feedback.

Points of light.

  1. Models and formations

    Choose models and thinking levels, configure teams, and run agent graphs.

  2. Repository tools

    Search files, inspect symbols, follow definitions, and review diffs.

  3. Content-checked edits

    Hashline editing checks file content before applying anchored changes.

  4. Programmable tools

    Run tool calls, loops, and filters locally in JavaScript with code_mode.

  5. Long-session context

    Compact, deduplicate, and age retained context as work continues.

  6. Project memory

    Review source-linked knowledge in Atlas and reuse it in later tasks.

  7. Persistent goals

    Resume sessions and autonomous work with configurable run limits.

  8. Headless runs

    Record task settings, source identity, tool activity, and acceptance evidence.

  9. Measured campaigns

    Evaluate isolated attempts with verifiers and independent review.

  10. Research loops

    Use Sloptomizer suggestions, Deli deliberation, and paired experiments.

  11. Measured benchmarks

    Calculate measured changes from paired benchmark samples.

  12. World TUI

    Introducing world tui foundations for reviewing work, building internal world models, presenting data graphs and agent behavior.

A coding and research harness.

angelX is a tool for explorers pushing the boundaries of research with AI. It is also a coding tool, a mini world, and a growing, self-reflective piece of software — built to evolve over time and adapt to the person driving it.

Under the hood: 200k+ lines of Rust focused on performance in autoresearch, agentic fleet control, and coding.

CUSTOM TUI LAYOUT
agent thinking · actions as primary surfaces
DEEP TOOLING
research · loops
WORLD MODEL
self-referential · adapts per user
FIELD PROVEN
kernel · cryptography research