ZYR3

ZYR3 is an agentic coding assistant that runs in your terminal: an interactive console, an autonomous build agent, and a lightweight local coding layer that completes code instantly on-device.

• Console — chat with ZYR3 in your terminal. Every message gets an instant local code-completion hint before the routed model answers.
• Build agent — point it at a repository and a task; it edits files, runs your test suite, and iterates on real feedback until the task is done.
• Local coding layer — a compact on-device model that generates code completions offline, with no GPU.

Install

pip install https://huggingface.co/zyr-AGENT/zyr3/resolve/main/cli/zyr3-build-0.4.0.tar.gz

For the interactive console (adds the Textual dependency):

pip install 'zyr3-build[tui]'

Quick start

# interactive console
zyr3 tui

# agentic coding loop in the current directory
zyr3 build "Fix the failing tests in src/"

# agentic loop in a specific workspace
zyr3 build "Refactor this repo" --dir ~/projects/app

# one-shot chat
zyr3 ask "Explain this traceback"

# offline code completion from the local layer (no network needed)
echo "def add(a, b):" | zyr3 layer

# defensive-security benchmark
zyr3 bench

# inspect lessons the agent has learned
zyr3 memory

# show effective configuration (secrets are masked)
zyr3 config

Console

zyr3 tui is a keyboard-driven terminal console with two tabs:

• Chat — send messages to the routed model. The local coding layer runs automatically on your prompt and its completion is appended as a hint, so you always get a fast first draft. A status line shows the live model pool size.
• Memory — view or clear the lessons the agent has learned from past runs.

Key bindings: q quit, Ctrl+Enter send, Ctrl+N / Ctrl+P next/prev tab, Ctrl+R refresh memory. Network calls and the local model run in worker threads, so the UI stays responsive.

Build agent

The build command works in rounds: it inspects the repo, writes complete files, runs shell commands (including your test suite), reads the real output, and fixes mistakes until the task is DONE (tests pass, exit 0). It never modifies test files and never emits diffs from imagination.

ZYR3 also learns as it works: after a failed round it records a short lesson, and relevant past lessons are injected into future prompts so the same mistake is not repeated. Manage this with zyr3 memory.

Configuration

All credentials and endpoints are read from the environment at runtime — a local gitignored .env.local next to the CLI is loaded automatically. Nothing is hardcoded in source; zyr3 config shows the effective settings with secrets masked. Every setting can also be overridden per command with a matching flag.

Benchmark

ZYR3 ships a defensive-security benchmark (two 100-point pillars, blue-team scope): attack analysis (detecting SQL injection, brute force, webshells, base64 payloads, phishing) and defense & hardening (parameterized SQL, safe subprocess handling, upload validation, SSH hardening). Run it with zyr3 bench.

Standard evals

The default heavy-tier model is also measured on hard public benchmarks with the official lm-evaluation-harness, the same harness used by public leaderboards:

Benchmark	Task	Score
MMLU-Pro	math	93.3%
MMLU-Pro	computer science	90.0%
MMLU-Pro	physics	100%
BIG-Bench Hard	logical deduction (7 objects)	50.0%

Each row is a 30-question sample with a fixed seed, greedy decoding, single pass. Scores are measured end-to-end through ZYR3's API gateway with the default heavy-tier model.

License

ZYR3's use license: zyr3-free-use-1.0 (see LICENSE).