angelX is a terminal coding agent. It works inside your repository with the models you choose, asks before it runs a command or edits a file, checks its work before it reports done, and keeps what it learns about your project.
curl -fsSL https://raw.githubusercontent.com/newjordan/angelX/main/install.sh | sh
cd your-project && angelXFor Linux x86_64 and arm64 (glibc 2.31 or newer) and macOS on Apple silicon.
The first launch finds the models you already have and connects one: Ollama,
LM Studio, llama.cpp or vLLM on this machine, an API key, or a ChatGPT or Grok
plan login. angelX setup changes it later, and angelX update installs the
latest release.
To build from source instead, you need Rust/Cargo, C/C++ tools, Bash, Node.js
and Python 3 (on macOS: xcode-select --install, then Rust from https://rustup.rs):
git clone https://github.com/newjordan/angelX.git
cd angelX
./bin/angelXModel setup · /commands · Feature evidence · Attributions · MIT
1 · Choose a model. /model lists every connected route with its thinking level; /think changes the level.
2 · Approve what it does. Commands and edits arrive as action capsules: approve one (y), approve the rest of the turn (a), or deny (n).
3 · Content-checked edits. Each edit is anchored to the file's current content and shows its byte change before it lands. /diff shows the result.
4 · Checked before done. The agent runs the checks and reports with receipts: every tool call, its time and the model route.
5 · Longer work. /goal sets a durable objective with acceptance criteria and a check. /loop runs it within time, iteration and token limits.
 Formations · model teams, from a lone coordinator to a full roster. |
 Context budget · where the window goes; compaction runs automatically. |
 Commands · /help and Tab completion for every command. |
 World view · Watch model behavior as an adventure in information. |
 Graph garden · every graph call grows a crop; a sprite carries each point. |
 Planted · /world crops reports the exact values it grew. |
- Models and formations — Choose models and thinking levels, configure teams, and run agent graphs.
- Repository tools — Search files, inspect symbols, follow definitions, and review diffs.
- Content-checked edits — Hashline editing checks file content before applying anchored changes.
- Programmable tools — Run tool calls, loops, and filters locally in JavaScript with
code_mode. - Long-session context — Compact, deduplicate, and age retained context as work continues.
- Project memory — Review source-linked knowledge in Atlas and reuse it in later tasks.
- Persistent goals — Resume sessions and autonomous work with configurable run limits.
- Headless runs — Record task settings, source identity, tool activity, and acceptance evidence.
- Measured campaigns — Evaluate isolated attempts with verifiers and independent review.
- Research loops — Use Sloptomizer suggestions, Deli deliberation, and paired experiments.
- Measured benchmarks — Calculate measured changes from paired benchmark samples.
- Book of behaviors — The harness steers the model with compact braille stamps; each stamp's English is taught the first time it appears.
- Adventure world model TUI — Introducing the early stages of Cyberdynamic world tui for reviewing work, presenting data graphs, adventure, and model behavior.
- Graph garden · 0.1.9 — Farmers and sprites turn actual
graphcalls into bar, line, and scatter crops on the fields. Visit with/world visit garden, inspect with/world crops, and begin again to regrow the bed. Guide and preview. - Calibrated seats · 0.1.9 — Muse Spark runs at minimal effort and Grok 4.7 at low, each picked on its hardest benchmark tasks, and playbook hints no longer send a model after a procedure its task rules out.
- Live research memory · 0.1.9 line — Sloptomizer connects ordinary checks and measured experiments through compact
⚠braille advice. Used legends return when the model changes or context is compacted. Ships in 0.1.92; ship notes. - A working adventure · 0.1.9 line — The knight follows real loop work through a 3D mine journey, with a School of Magic, study, and underground archive reflecting actual research evidence.
- The Delve · 0.1.92 —
/dungeonwalks your knight to the Delve's gate: a co-op knights-vs-monsters dungeon drawn in the realm's own pixel art. Friends join from their own angelX with/dungeon join, and cards, spells and wishes load while you play. Guide. - Cartridges · 0.1.93 — Plug a competition into the loop as a folder: a
cartridge.toml, and Rust when it needs more. Yours stay outside the tree. Guide and figures.
136 repository-repair tasks (48 JS, 34 Python, 30 Rust, 24 C++) from the Aider polyglot set. One attempt per task, 600 s limit, graded by each task's tests. Wall: median agent time per attempt. Tokens: totals per cell.
| model | solved | wall (median) | calls / task | input tokens | cache hit | output tokens |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash, thinking off · run 1 | 135 / 136 | 9.9 s | 6.0 | 8.7 M | 84% | 205 k |
| DeepSeek V4.1 Flash, thinking off · run 2 | 135 / 136 | 9.1 s | 5.3 | 6.9 M | 80% | 195 k |
Evaluator: Prime Intellect Verifiers v0.3.1 · angelX 0.1.8 · temperature 0 · 600 s per attempt · fresh environment per attempt · 2026-09-30. The other models' latest results are 0.1.6's, below.
| model | solved | wall (median) | calls / task | input tokens | cache hit | output tokens |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash, thinking off | 134 / 136 | 10.7 s | 7.8 | 12.7 M | 89% | 310 k |
| GLM-5.3-Flash, thinking low | 134 / 136 | 46.1 s | 7.9 | 9.4 M | 83% | 222 k |
| Grok 4.7, thinking low | 136 / 136 | 17.5 s | 5.0 | 6.7 M | 57% | 134 k |
| gpt-6-luna, thinking medium | 136 / 136 | 31.3 s | — | — | — | — |
| Muse Spark, thinking low | 136 / 136 | 25.5 s | 4.9 | 5.9 M | 54% | 311 k |
Evaluator: Prime Intellect Verifiers v0.3.1 · angelX 0.1.6 · temperature 0; Muse Spark 1.0 (Meta's recommended setting); gpt-6-luna on its ChatGPT plan, which reports no token counts · 600 s per attempt · fresh environment per attempt · 2026-09-23
| model | harness | solved | wall (median) | calls / task | input tokens | cache hit | output tokens |
|---|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash, thinking off | angelX | 133 / 136 | 9.9 s | 7.5 | 11.5 M | 88% | 257 k |
| OpenCode 1.18.31 | 132 / 136 | 8.6 s | 12.5 | 67.7 M | 97% | 349 k | |
| oh-my-pi 18.2.4 | 86 / 93 * | 15.3 s | 35.5 | 200.1 M | 99% | 866 k | |
| GLM-5.3-Flash, thinking low | angelX | 135 / 136 | 50.7 s | 8.9 | 10.9 M | 84% | 256 k |
| OpenCode 1.18.31 | 133 / 136 | 38.9 s | 7.4 | 10.1 M | 85% | 179 k | |
| oh-my-pi 18.2.4 | 133 / 136 | 37.7 s | 8.9 | 24.1 M | 90% | 203 k |
- oh-my-pi on DeepSeek reached its 200M-token budget cap after 93 tasks.
The full GLM cohort took 34.5% longer than OpenCode in the September 21 comparison and 17.6% longer in the September 23 Angel run. The later gap is in model time; Angel's measured non-model time was lower. Verification alone does not explain the difference. See the full 136-task slowdown audit for totals, tails, task-level causes and comparison limits.
Evaluator: Prime Intellect Verifiers v0.3.1 · angelX 98d7340 · OpenCode 1.18.31 · oh-my-pi 18.2.4 · DeepSeek V4.1 Flash, thinking off · GLM-5.3-Flash, thinking low (lowest available) · temperature 0 · 8,192-token output cap · 600 s per attempt · fresh environment per attempt · 2026-09-21
Research and public work that informed Angel:
- A Programming Paradigm for Spatiotemporal Composability — Yifan Shi, Wei Zhang and Tianyi Cui (2026); component lifecycle, reactive dependencies and reversible registration.
- GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning — Lakshya A. Agrawal and collaborators (2025); reflection and Pareto selection.
- Learning, Fast and Slow: Towards LLMs That Adapt Continually — Rishabh Tiwari and collaborators (2026); Sloptomizer's fast-context and slow-learner design.
- Finite-time Analysis of the Multiarmed Bandit Problem — Peter Auer, Nicolò Cesa-Bianchi and Paul Fischer (2002); UCB1 exploration.
- Prime Agent: A Self-Improving RLM Harness — Seth Karten and collaborators (2026); editable, persistent harness state.
- Geoffrey Huntley's Ralph loop — an agent put on a loop with a simple prime directive: read the issues, pick one, write the patch, run the tests, repeat. The measured campaign and autonomous loop work here is a direct descendant of that idea.
- RL Systems Mind the Gap: Matching Trainer and Generator Throughput — Kimbo Chen, Cheang Kang Wen and Dylan Patel / SemiAnalysis (2026); rollout staleness, pruning and reward-variance guards.
- Deli_AutoResearch — Deli Chen; retained findings, fresh-context deliberation and stall recovery.
- OpenScience — Synthetic Sciences; literature-search design. Search metadata comes from OpenAlex, Crossref, Semantic Scholar and Europe PMC.
Code and tooling credits include OpenAI Codex, Grok CLI, oh-my-pi, DeepSeek-Reasonix, Hermes Agent, Prime Agent, Dotmax, ureq, Ratatui, Crossterm, rusty_v8 and V8. Early inspiration: DotAgents and SpeakMCP by aj47. File-tool interface references include Claude Code, aider and OpenHands. Attributions records implementation links, authors and retained licenses; research inspiration and incorporated code are identified separately.
Field proven: 100+ records on Yukon public research leaderboards, with first-place results in kernel optimization, LLM inference and cryptography research, and 2nd place on the GPU MODE Cholesky leaderboard. https://www.yukon.org/